Web, Artificial Intelligence and Network Applications

Advances in Intelligent Systems and Computing 927
Leonard Barolli
Makoto Takizawa
Fatos Xhafa
Tomoya Enokido Editors
Web, Artificial
Intelligence
and Network
Applications
Proceedings of the Workshops of the
33rd International Conference on
Advanced Information Networking
and Applications (WAINA-2019)
Advances in Intelligent Systems and Computing
Volume 927
Series editor
Janusz Kacprzyk, Systems Research Institute, Polish Academy of Sciences,
Warsaw, Poland
e-mail: kacprzyk@ibspan.waw.pl
The series “Advances in Intelligent Systems and Computing” contains publications on theory,
applications, and design methods of Intelligent Systems and Intelligent Computing. Virtually all
disciplines such as engineering, natural sciences, computer and information science, ICT, economics,
business, e-commerce, environment, healthcare, life science are covered. The list of topics spans all the
areas of modern intelligent systems and computing such as: computational intelligence, soft computing
including neural networks, fuzzy systems, evolutionary computing and the fusion of these paradigms,
social intelligence, ambient intelligence, computational neuroscience, artificial life, virtual worlds and
society, cognitive science and systems, Perception and Vision, DNA and immune based systems,
self-organizing and adaptive systems, e-Learning and teaching, human-centered and human-centric
computing, recommender systems, intelligent control, robotics and mechatronics including
human-machine teaming, knowledge-based paradigms, learning paradigms, machine ethics, intelligent
data analysis, knowledge management, intelligent agents, intelligent decision making and support,
intelligent network security, trust management, interactive entertainment, Web intelligence and multimedia.
The publications within “Advances in Intelligent Systems and Computing” are primarily proceedings
of important conferences, symposia and congresses. They cover significant recent developments in the
field, both of a foundational and applicable character. An important characteristic feature of the series is
the short publication time and world-wide distribution. This permits a rapid and broad dissemination of
research results.
** Indexing: The books of this series are submitted to ISI Proceedings, EI-Compendex, DBLP,
SCOPUS, Google Scholar and Springerlink **
Advisory Board
Chairman
Nikhil R. Pal, Indian Statistical Institute, Kolkata, India
e-mail: nikhil@isical.ac.in
Members
Rafael Bello Perez, Faculty of Mathematics, Physics and Computing, Universidad Central de Las Villas, Santa
Clara, Cuba
e-mail: rbellop@uclv.edu.cu
Emilio S. Corchado, University of Salamanca, Salamanca, Spain
e-mail: escorchado@usal.es
Hani Hagras, School of Computer Science & Electronic Engineering, University of Essex, Colchester, UK
e-mail: hani@essex.ac.uk
László T. Kóczy, Department of Information Technology, Faculty of Engineering Sciences, Győr, Hungary
e-mail: koczy@sze.hu
Vladik Kreinovich, Department of Computer Science, University of Texas at El Paso, El Paso, TX, USA
e-mail: vladik@utep.edu
Chin-Teng Lin, Department of Electrical Engineering, National Chiao Tung University, Hsinchu, Taiwan
e-mail: ctlin@mail.nctu.edu.tw
Jie Lu, Faculty of Engineering and Information, University of Technology Sydney, Sydney, NSW, Australia
e-mail: Jie.Lu@uts.edu.au
Patricia Melin, Graduate Program of Computer Science, Tijuana Institute of Technology, Tijuana, Mexico
e-mail: epmelin@hafsamx.org
Nadia Nedjah, Department of Electronics Engineering, University of Rio de Janeiro, Rio de Janeiro, Brazil
e-mail: nadia@eng.uerj.br
Ngoc Thanh Nguyen, Wrocław University of Technology, Wrocław, Poland
e-mail: Ngoc-Thanh.Nguyen@pwr.edu.pl
Jun Wang, Department of Mechanical and Automation, The Chinese University of Hong Kong, Shatin,
Hong Kong
e-mail: jwang@mae.cuhk.edu.hk
More information about this series at http://www.springer.com/series/11156

Leonard Barolli Makoto Takizawa
• •
Fatos Xhafa Tomoya Enokido

•
Editors
Web, Artificial Intelligence

and Network Applications
Proceedings of the Workshops of the 33rd
International Conference on Advanced
Information Networking and Applications
(WAINA-2019)
123
Editors
Leonard Barolli Makoto Takizawa
Department of Information Department of Advanced Sciences
and Communication Engineering Hosei University
Fukuoka Institute of Technology Koganei-Shi, Tokyo, Japan
Fukuoka, Japan
Fatos Xhafa Tomoya Enokido

Department of Computer Science Faculty of Business Administration
Technical University of Catalonia Rissho University
Barcelona, Barcelona, Spain Tokyo, Japan
ISSN 2194-5357 ISSN 2194-5365 (electronic)

Advances in Intelligent Systems and Computing
ISBN 978-3-030-15034-1 ISBN 978-3-030-15035-8 (eBook)
https://doi.org/10.1007/978-3-030-15035-8
Library of Congress Control Number: 2019933333
© Springer Nature Switzerland AG 2019

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part
of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission
or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar
methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt from
the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authors or the editors give a warranty, expressed or implied, with respect to the material contained
herein or for any errors or omissions that may have been made. The publisher remains neutral with regard
to jurisdictional claims in published maps and institutional affiliations.
This Springer imprint is published by the registered company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Welcome Message from AINA-2019 Steering
Committee Co-chairs
Welcome to the 33rd International Conference on Advanced Information

Networking and Applications (AINA-2019). It is our great pleasure and honor to
held AINA-2019 at Kunibiki Messe, Matsue, Japan, from March 27–29, 2019. On
behalf of the AINA Steering Committee and AINA-2019 Organizing Committee,
we would like to express to all participants our cordial welcome and high respect.
AINA is an international forum, where scientists and researchers from academia
and industry working in various scientific and technical areas of networking and
distributed computing systems can demonstrate new ideas and solutions in
distributed computing systems.
AINA was born in Asia, but it is now an international conference with high
quality thanks to the great help and cooperation of many international friendly
volunteers. AINA is a very open society and is always welcoming international
volunteers from any country and any area in the world. In conjunction with
AINA-2019 main conference, we have also 14 international workshops.
An international conference can be organized by support and great voluntary
efforts of many people and organizations. Our main responsibility is to coordinate
various tasks carried out with other willing and talented volunteers.
We would like to thank AINA-2019 General Co-chairs, PC Co-chairs,
Workshop Co-chairs, Track Area Chairs, PC Members, and Workshop Organizers
for their great efforts to make AINA-2019 a very successful event. We have special
thanks to the Finance Chair and Web Administrator Co-chairs.
We would like to take opportunity to thank all members of the Organization
Committee and Program Committee as well as all Reviewers for their hard work to
make the reviews on time and authors for submitting the papers. We would like to
thank Local Arrangement Team for the technical support and good local arrange-
ment for the conference.
v
vi Welcome Message from AINA-2019 Steering Committee Co-chairs
Finally, we would like to thank Matsue City, Shimane Prefecture, The

Telecommunications Advancement Foundation (TAF), Japan, for their financial
support.
We do hope that you will have a great time in Matsue, Japan.
Makoto Takizawa
Leonard Barolli
AINA Steering Committee Co-chairs
Welcome Message from AINA-2019 General
Co-chairs
It is our great pleasure to welcome you all at the 33rd International Conference on
Advanced Information Networking and Applications (AINA-2019), which will be
held at Kunibiki Messe, Matsue, Japan, from March 27–29, 2019.
AINA International Conference is a forum for sharing ideas and research work in
the emerging areas of information networking and their applications. The area of
advanced networking has grown very rapidly, and the applications around it have
experienced an explosive growth especially in the area of pervasive and mobile
applications, sensor networks, ad hoc networks, vehicular networks, multimedia
computing and social networking, semantic collaborative systems, as well as grid,
P2P, IoT, and cloud computing. This advanced networking revolution is trans-
forming the way people live, work, and interact with each other, and is impacting
the way business, education, entertainment, and health care are operating. The
papers included in the proceedings cover theory, design, and application of
computer networks, distributed computing, and information systems.
Each year AINA receives a lot of paper submissions from all around the world. It
has maintained high-quality accepted papers and is aspiring to be one of the main
international conferences on the information networking in the world. In
conjunction with AINA-2019 conference, there are 14 workshops, which also
accepted good-quality papers.
An international conference of this size requires the support and help of many
people. A lot of people have helped and worked hard to produce a successful
AINA-2019 technical program and conference proceedings. First, we would like to
thank all authors for submitting their papers, the session chairs, and distinguished
keynote speakers. We are indebted to program area chairs, Program Committee
Members and Reviewers, who carried out the most difficult work of carefully
evaluating the submitted papers.
We would like to give our special thanks to Prof. Makoto Takizawa and Prof.
Leonard Barolli, the Co-chairs of the Steering Committee for their strong encour-
agement, guidance, and insights, and for spending a lot of energy for conference
organization and shaping the conference program. We would like to thank PC
Co-chairs and Workshop Co-chairs of AINA-2019 for their great contribution to the
vii
viii Welcome Message from AINA-2019 General Co-chairs
success of the conference. Our special thanks go to the Finance Chair and Web
Administrator Co-chairs.
Finally, we would like to thank the Local Arrangement Team for the support and
good local arrangement for the conference. We do hope that you will have a great
time in Matsue, Japan.
Tomoya Enokido
Farookh Hussain
Alireza Shahrabi
AINA-2019 General Co-chairs
Welcome Message from AINA-2019 Program
Committee Co-chairs
Welcome to the 33rd International Conference on Advanced Information

Networking and Applications (AINA-2019), which will be held at Kunibiki Messe,
Matsue, Japan, from March 27–29, 2019.
The purpose of AINA conference is to bring together researchers, developers,
and industrial experts to share new ideas and recent research results in the emerging
areas of information networking and their applications. The papers included in the
proceedings cover all aspects of theory, design, and application of computer net-
works and distributed computing systems. Most of the papers deal with new trends
in information networking, such as wireless sensor networks, ad hoc networks,
cloud computing, peer-to-peer systems, grid computing, pervasive and ubiquitous
systems, multimedia systems, security, multi-agent systems, IoT, and Web-based
systems.
This edition AINA received many paper submissions from all over the world.
Each submission was peer-reviewed by Program Committee members and invited
external Reviewers. Finally, the Program Committee accepted 112 papers (about
25% acceptance ratio), which will be presented during the conference days.
Unfortunately, many interesting and good papers could not be accepted in
AINA-2019 due to the limited number of time slots allocated for presentations at
the conference.
We are very proud and honored to have two distinguished keynote talks by
Dr. Markus Aleksy, ABB AG, Germany, and Naohiro Hayashibara, Kyoto Sangyo
University, Japan, who will present their recent work and will give new insights and
ideas to the conference participants.
Organizing an international conference of this size is of course a team effort.
Therefore, we gladly admit that we had the help of many very professional people.
First of all, we would like to thank all the authors for their interesting contributions
since they shape the program and make it interesting for the audience. Moreover,
we would like to express our thankfulness to all program vice-chairs for their great
efforts. Additionally, we would like to thank all Program Committee members and
Reviewers who carried out the most important work to evaluate the submitted
papers. We also thank the Workshop Co-chairs for organizing many excellent
ix
x Welcome Message from AINA-2019 Program Committee Co-chairs
workshops and symposiums, which enrich the conference and provide additional
opportunities for discussions and future cooperations.
The great success of the AINA conference series would not be possible without
the enormous commitment and support of the Steering Committee co-chairs Prof.
Makoto Takizawa and Prof. Leonard Barolli. Therefore, we would like to thank
them for their strong encouragement and guidance.
The general coordination of an event such as AINA conference requires a lot of
coordination effort as well as many other activities related to the conference
organization. Here, we thank the General Co-chairs for their great support and
invaluable suggestions. We give special thanks to the Finance Chair and Web
Administrator Co-chairs for their great efforts and efficient work to deal with many
conference matters.
We hope you will enjoy the conference and readings and have a great time in
Matsue, Japan.
Akimitsu Kanzaki
Flora Amato
Omar Hussain
AINA-2019 Program Committee Co-chairs
Welcome Message from AINA-2019 Workshops’
Co-chairs
Welcome to AINA-2019 Workshops to be held in conjunction with the 33rd

International Conference on Advanced Information Networking and Applications
(AINA-2019) at Kunibiki Messe, Matsue, Japan, from March 27–29, 2019. The
goal of AINA workshops is to provide a forum for international researchers and
practitioners to exchange and share their new ideas, research results, and ongoing
work on leading-edge topics in the different fields of information networks and their
applications. Some of the accepted workshops deal with topics that open up per-
spectives beyond the ordinary, thus, enriching the topics usually addressed by the
AINA conference.
For this edition, the following 14 symposiums and workshops will be held with
AINA-2019.
1. The 15th International Symposium on Frontiers of Information Systems and
Network Applications (FINA-2019)
2. The 15th International Workshop on Heterogeneous Wireless Networks
(HWISE-2019)
3. The 12th International Symposium on Mining and Web (MAW-2019)
4. The 12th International Workshop on Bio and Intelligent Computing
(BICom-2019)
5. The 12th International Workshop on Telecommunication Networking,
Applications and Systems (TeNAS-2019)
6. The 10th International Workshop on Disaster and Emergency Information
Network Systems (IWDENS-2019)
7. The 7th International Workshop on Collaborative Emerging Systems
(COLLABES-2019)
8. The 6th International Workshop on Security Intricacies in Cyber-Physical
Systems and Services (INTRICATE-SEC-2019)
9. The 5th International Workshop on Engineering Energy Efficient Internet
Worked Smart seNsors (E3WSN-2019)
10. The 4th International Workshop on Innovative Technologies in Informatics and
Networking (WITIN-2019)
xi
xii Welcome Message from AINA-2019 Workshops’ Co-chairs
11. The 4th International Workshop on Big Data Processing in Online Social
Network (BOSON-2019)
12. The 2nd International Workshop on Internet of Everything and Machine
Learning Applications (IOEMLA-2019)
13. The 1st International Workshop on Multi-Clouds and Mobile Edge Computing
(M2EC-2019)
14. The 1st Workshop on Artificial Intelligence and Machine Learning
(AIMAL-2019)
We would like to thank the community for their great response to AINA-2019
workshops. The excellent technical program of the workshops was the result of a
professional work from Workshop Chairs, workshop Program Committees,
Reviewers, and authors.
We would like to give our special thanks to Prof. Makoto Takizawa and
Prof. Leonard Barolli, the Steering Committee Chairs of AINA International
Conference, for their strong encouragement and guidance to organize the
AINA-2019 workshops and symposiums. We would like to thank AINA-2019
General Co-chairs their advices to make possible organization of AINA-2019
workshops and symposiums. We are thankful to AINA-2019 Program Co-chairs for
their support and help to prepare the technical program of AINA-2019 workshops
and symposiums.
We wish all of you entertaining and rewarding experience in all workshops and
AINA-2019 International Conference.
Hui-Huang Hsu
Omid Ameri Sianaki
Rubem Pereira
AINA-2019 Workshop Co-chairs
Welcome Message from AINA-2019 Workshops’ Co-chairs xiii
AINA-2019 Organizing Committee
General Co-chairs
Tomoya Enokido Rissho University, Japan

Farookh Hussain University of Technology, Sydney, Australia
Alireza Shahrabi Glasgow Caledonian University, UK
Program Committee Co-chairs
Akimitsu Kanzaki Shimane University, Japan

Flora Amato University of Naples Federico II, Italy
Omar Hussain University of New South Wales, Australia
Workshop Co-chairs
Hui-Huang Hsu Tamkang University, Taiwan

Omid Ameri Sianaki Victoria University, Australia
Rubem Pereira Liverpool John Moores University, UK
International Special Issue Journal Co-chairs
Fatos Xhafa Technical University of Catalonia, Spain

David Taniar Monash University, Australia
Isaac Woungang Ryerson University, Canada
Award Co-chairs
Marek Ogiela AGH University of Science and Technology, Poland

Kin Fun Li University of Victoria, Canada
Markus Aleksy ABB AG, Germany
Fang-Yie Leu Tunghai University, Taiwan
Publicity Co-chairs
Arjan Durresi IUPUI, USA

Akio Koyama Yamagata University, Japan
Wenny Rahayu La Trobe University, Australia
Lidia Ogiela AGH University of Science and Technology, Poland
xiv Welcome Message from AINA-2019 Workshops’ Co-chairs
International Liaison Co-chairs
Nadeem Javaid COMSATS University Islamabad, Pakistan

Minoru Uehara Toyo University, Japan
Hsing-Chung Chen Asia University, Taiwan
Local Arrangement Co-chairs
Elis Kulla Okayama University of Science, Japan

Keita Matsuo Fukuoka Institute of Technology, Japan
Finance Chair
Makoto Ikeda Fukuoka Institute of Technology, Japan
Web Chairs
Donald Elmazi Fukuoka Institute of Technology, Japan

Yi Liu Fukuoka Institute of Technology, Japan
Miralda Cuka Fukuoka Institute of Technology, Japan
Kevin Bylykbashi Fukuoka Institute of Technology, Japan
Steering Committee Chairs
Makoto Takizawa Hosei University, Japan

Leonard Barolli Fukuoka Institute of Technology, Japan
Welcome Message from FINA-2019 Symposium
Co-chairs
Welcome to the 15th International Symposium on Frontiers of Information Systems

and Network Applications (FINA-2019), which will be held in conjunction with the
33rd International Conference on Advanced Information Networking and
Applications (AINA-2019) at Kunibiki Messe, Matsue, Japan, from March 27–29,
2019.
The purpose of FINA symposium is to bring together developers and researchers
to share ideas and research work in the emerging areas of information systems and
their applications. Information systems and networks of today are going through a
rapid evolution and are giving a great influence to our society. In addition,
computing models are changing to new architecture like cloud computing, grid
computing, and peer-to-peer (P2P) computing models. For this reason, it is sig-
nificant to discuss how to design and realize high-performance computation models
and network applications for distributed systems. The papers included in the pro-
ceedings cover novel theories, designs, and applications of information systems and
network applications in advanced distributed system environments.
FINA-2019 contains high-quality research papers submitted by researchers from
all over the world. Each submitted paper was peer-reviewed by Reviewers who are
experts in the subject area of the paper. Based on the review results, the Program
Committee accepted good-quality papers.
For organizing an international symposium, it is needed the support and help of
many people. First, we would like to thank all authors for submitting their papers.
We also appreciate the support from Program Committee members and Reviewers
who carried out the most difficult work of carefully evaluating the submitted papers.
We would like to give our special thanks to Prof. Makoto Takizawa and Prof.
Leonard Barolli, the Steering Committee Chairs of AINA, for their strong
encouragement and guidance to organize the symposium. We would like to thank
AINA-2019 General Co-chairs for their advices to make possible organization of
FINA-2019. We also would like to thank AINA-2019 Workshop Co-chairs for their
support.
xv
xvi Welcome Message from FINA-2019 Symposium Co-chairs
We wish all of you entertaining and rewarding experience in FINA-2019

symposium and AINA-2019 conference.
Akimitsu Kanzaki
Flora Amato
Omar Hussain
FINA-2019 Symposium Co-chairs
Welcome Message from FINA-2019 Symposium Co-chairs xvii
FINA-2019 Organizing Committee
FINA-2019 Symposium Co-chairs
Akimitsu Kanzaki Shimane University, Japan

Flora Amato University of Naples Federico II, Italy
Omar Hussain University of New South Wales, Australia
Program Committee Members

Elis Kulla Okayama University of Science, Japan
Marek R. Ogiela AGH University of Science and Technology, Poland
Markus Aleksy ABB AG, Germany
Muhammad Younas Oxford Brookes University, UK
Arjan Durresi IUPUI, USA
Farookh Hussain University of Technology, Sydney, Australia
Fang-Yie Leu Tunghai University, Taiwan
Welcome Message from HWISE-2019
International Workshop Co-chairs
Welcome to the 15th International Workshop on Heterogeneous Wireless Networks

(HWISE-2019), which will be held in conjunction with the 33rd International
Conference on Advanced Information Networking and Applications (AINA-2019)
at Kunibiki Messe, Matsue, Japan, from March 27–29, 2019.
The goal of HWISE International Workshop is to bring together computer sci-
entists, industrial engineers, and researchers to discuss and exchange experimental
or theoretical results, novel designs, work-in-progress, experience, case studies, and
trendsetting ideas in the area of heterogeneous wireless networks.
This international workshop collected research papers on the above research
issues from all over the world. Papers collected in this international workshop were
carefully reviewed, and according to the review results, the Program Committee
members selected high-quality papers.
Many people have kindly helped us prepare and organize the HWISE-2019
workshop. First, we would like to thank the authors who submitted the papers and
Reviewers who carefully evaluated the submitted papers. We would like to give our
special thanks to Prof. Makoto Takizawa, as Steering Committee Co-chair of
AINA, for his strong encouragement and guidance to organize the workshop. We
would like to thank especially AINA-2019 General Co-chairs for their advices and
support to make possible organization of HWISE-2019. We also, would like to
thank Program Committee Co-chairs and Workshop Co-chairs for their help and
directions.
We hope you will enjoy the conference and have a great time in Matsue, Japan.
Leonard Barolli
Arjan Durresi
HWISE-2019 International Workshop Co-chairs
xix
xx Welcome Message from HWISE-2019 International Workshop Co-chairs
HWISE-2019 Organizing Committee
Workshop Co-chairs

Arjan Durresi Indiana University Purdue University at Indianapolis,
USA
Advisory Co-chairs

Raj Jain Washington University in St. Louis, USA

Sriram Chellappan University of South Florida, USA
Admir Barolli Aleksander Moisiu University of Durres, Albania
Tomoya Enokido Tokyo Denki University, Japan
Mukul Goyal University of Wisconsin Milwaukee, USA
Sanjay Kumar Madria University of Missouri-Rolla, USA
Masakatsu Morii Kobe University, Japan
Masakatsu Nishigaki Shizuoka University, Japan
Vamsi Paruchuri University of Central Arkansas, USA
Yoshitaka Shibata Iwate Prefectural University, Japan
Kaoru Sugita Fukuoka Institute of Technology, Japan
Fatos Xhafa Polytechnic University of Catalonia, Spain
Eric Pardede La Trobe University, Australia
Muhammad Younas Oxford Brookes University, UK
Mimoza Durresi European University of Tirana, Albania
Web Administrator Co-chairs

Welcome Message from MAW-2019
International Symposium Co-chairs
Welcome to the 12th International Symposium on Mining and Web (MAW-2019),

which will be held in conjunction with the 33rd International Conference on
Advanced Information Networking and Applications (AINA-2019) at Kunibiki
Messe, Matsue, Japan, from March 27–29, 2019.
As the Web has become a major source of information, techniques and
methodologies to extract quality information are of paramount importance for many
Web applications and users. Data mining and knowledge discovery play key roles
in many of today’s prominent Web applications such as e-commerce and computer
security. MAW-2019 symposium aims to bring together scientists, engineers, and
practitioners to discuss, exchange ideas, and present their research on mining the
Web.
After a thorough reviewing process conducted by the Program Committee, we
selected high-quality papers for presentation at the symposium and publication in
the proceedings.
We would like to express our sincere gratitude to the members of the Program
Committee for their efforts and to AINA-2019 Organizing Committee for
co-hosting MAW-2019. Most importantly, we thank all the authors for their sub-
mission and contribution to the symposium.
Kin Fun Li
Kosuke Takano
MAW-2019 Symposium Co-chairs
xxi
xxii Welcome Message from MAW-2019 International Symposium Co-chairs
MAW-2019 Organizing Committee

Symposium Organizers

Kosuke Takano Kanagawa Institute of Technology, Japan
Takahiro Hara Osaka University, Japan

Hayato Yamana Waseda University, Japan
Gulnar Ali Xinjiang University, China
Deepali Arora University of Victoria, Canada
You Chen Vanderbilt University, USA
Belkacem Chikhaoui University of Sherbrooke, Canada
Tomohiro Fukuhara University of Tokyo, Japan
Michael Horie University of Victoria, Canada
Wei Lu Keene University, USA
Kotaro Nakayama University of Tokyo, Japan
Ana-Maria Sevcenco University of Victoria, Canada
Martine Wedlake IBM, USA
Wei Wei Xian University of Technology, China
Jinmin Yang Hunan University, China
Shengrui Wang University of Sherbrooke, Canada
Welcome Message from BICom-2019
Welcome to the 12th International Workshop on Bio and Intelligent Computing

(BICom-2019), which will be held in conjunction with the 33rd International
The workshop is intended to facilitate exchange of ideas and collaborations
between researchers form computer science, biology, medicine, mathematics,
statistics, physics, intelligent computing, and such related sciences, to discuss
various aspects of bio-computing, intelligent computing, and their applications.
This international workshop collected research papers on the above research
issues from all over the world. Papers collected in this international workshop were
carefully reviewed. According to the review results, the Program Committee
members selected high-quality papers to be presented in this workshop.
Many people have kindly helped us to prepare and organize the BICom-2019
workshop. First, we would like to thank the authors who submitted high-quality
papers and Reviewers who carefully evaluated the submitted papers. We would like
to give our special thanks to Prof. Makoto Takizawa, as Steering Committee
Co-chair of AINA-2019, for his strong encouragement and guidance to organize the
workshop. We would like to thank especially AINA-2019 General Co-chairs, PC
Co-chairs, and Workshop Co-chairs for their advices and support to make possible
organization of BICom-2019.
Leonard Barolli
Arjan Durresi
BICom-2019 Co-chairs
xxiii
xxiv Welcome Message from BICom-2019 International Workshop Co-chairs
BICom-2019 Organizing Committee
Workshop Co-chairs

Arjan Durresi Indiana University Purdue University
at Indianapolis, USA
Advisory Co-chairs

Raj Jain Washington University in St. Louis, USA
Mimoza Durresi European University of Tirana, Albania

Jake Chen Indiana University Purdue University at Indianapolis,
USA
Tomoya Enokido Tokyo Denki University, Japan
Mukul Goyal University of Wisconsin Milwaukee, USA
Admir Barolli Aleksander Moisiu University of Durres, Albania
Chunlei Liu Troy University, USA
Arben Markoci Institut CatalC de Nanotecnologia Barcelona, Spain
Snehasis Indiana University Purdue University Indianapolis,
Mukhopadhyay USA
Mathew Palakal Indiana University Purdue University Indianapolis,
USA
Fatos Xhafa Polytechnic University of Catalonia, Spain
Wenye Wang North Carolina State University, USA
Hsiao-Chun Wu Louisiana State University, USA
Mohamed Younis University of Maryland, USA
Web Administrator Co-chairs

Welcome Message from TeNAS-2019 Workshop
Organizers
We wish to extend a warm welcome to everyone attending the 13th International

Workshop on Telecommunication Networking, Applications and Systems
(TeNAS-2019), which is being held in conjunction with the 33rd International
in Matsue, Japan. We would like to thank AINA-2019 Organizing Committee for
providing us with this forum to bring together researchers and practitioners from
across the world.
For the TeNAS-2019 workshop, we received submissions from various countries
including academic and industrial contributions. Submissions were reviewed by
two referees. The Program Committee, which consisted of researchers and practi-
tioners from different fields in networking, security, applications, and distributed
systems, selected good-quality papers to be presented in the workshop, covering a
wide range of related topics.
Many people contributed to the organization of TeNAS-2019. We wish to thank
the Program Committee members, the Steering Committee members, and the
Reviewers for their great work.
Rubem Pereira
Jeu-Yih Jeng
TeNAS-2019 General Chairs
Teh-Sheng Huang
Shin-Hao Chang
TeNAS-2019 Program Committee Co-chairs
xxv
xxvi Welcome Message from TeNAS-2019 Workshop Organizers
TeNAS-2019 Organizing Committee
Steering Committee Chair
Timothy K. Shih National Central University, Taiwan
General Co-chairs
Rubem Pereira Liverpool John Moores University, UK

Jeu-Yin Jeng ChungHwa Telecom, Taiwan
Program Co-chairs
Teh-Sheng Huang ChungHwa Telecom, Taiwan

Shih-Hao Chang Tamkang, Taiwan
Huang Min-Ku ChungHwa Telecom, Taiwan

Pi-Chung Wang National Chun-Hsing University, Taiwan
Benxiong Huang Huazhong University, China
Irfan Awan University of Bradford, UK
Tan-Hsu Tan National Taipei University of Technology, Taiwan
Tsai Chung-ping Long Island University, USA
Ming-Chi Lee National Ping-Tung Institute, Taiwan
Geyong Min University of Exeter, UK
Sheng-Ping Bill L. Cisco Networks, USA
Chelsea Dobbins Liverpool John Moores University, UK
John Haggerty Nottingham Trent University, UK
Ho-Jin Choi KAIST, Korea
Karim Djemame University of Leeds, UK
Ella Pereira Edge Hill University, UK
Lawrence Deng St John’s University, Taiwan
Paul Fergus Liverpool John Moores University, UK
Lu Fang-Sun ChungHwa Telecom, Taiwan
Junyang Zhou Baptist University, Hong Kong
Rob Hegarty Manchester Metropolitan University, UK
Stefan Wallin Data Ductus, Sweden
Welcome Message from TeNAS-2019 Workshop Organizers xxvii
Teh-Lung Liu NCHC, Taiwan

William Hurst Liverpool John Moores University, UK
Hui Cheng Liverpool John Moores University, UK
Gyu Myoung Lee Liverpool John Moores University, UK
Welcome Message from IWDENS-2019
Workshop Organizers
Welcome to the 10th International Workshop on Disaster and Emergency

Information Network Systems (IWDENS-2019), which will be held in conjunction
with the 33rd International Conference on Advanced Information Networking and
2019.
This workshop serves as a forum for the exchange of information and ideas
regarding with not only network design and implementation and application sys-
tems for large-scale national disasters such as earthquake, mountain explosion,
Tsunami in addition to ordinal disasters, such as typhoon, rain flooding, and
snowslide. Not only technological aspects by researchers but also actual experience
by governmental officers or volunteers for disasters are also discussed.
A massive 9.1 magnitude earthquake hit over the northern Japan on March 11,
2011, and there were a number of papers submitted as the reports and researches
of the earthquake. We hope that all participants of this workshop can share ideas
and research works in the emerging areas of Disaster and Emergency Information
Network and Systems.
Many people have kindly helped us to prepare and organize the IWDENS
workshop. First of all, we would like to thank AINA-2019 Organization Committee
for their support, guidance, and help for making the workshop. We would like to
express our special thanks to all of IWDENS Program Committee members and
Reviewers for organizing the workshop and reviewing the submitted papers,
respectively. We also appreciate the member of recovery activities of the Great East
Japan Earthquake and Tsunami for cooperating with this workshop.
Finally, we would like to give our special thanks to all members of Regional
Disaster Information Network Research Institute in Iwate, Japan, for supporting and
promoting IWDENS-2019.
Yoshitaka Shibata
Noriki Uchida
IWDENS-2019 Workshop Co-chairs
xxix
xxx Welcome Message from IWDENS-2019 Workshop Organizers
IWDENS-2019 Organizing Committee

Program Co-chairs
Yoshitaka Shibata Iwate Prefectural University, Japan

Noriki Uchida Fukuoka Institute of Technology, Japan
Hiroaki Yuze University of Shizuoka, Japan

Kazuo Takahata Saitama Institute of Technology, Japan
Go Hirakawa Network Applied Laboratory, Japan
Jun Sawamoto Iwate Prefectural University, Japan
Kazuhiko Nakamura National Institute of Information and Communication
Technology (NICT), Japan
Tomoyuki Ishida Fukuoka Institute of Technology, Japan
Welcome Message from COLLABES-2019
International Workshop Organizer
Welcome to the 7th International Workshop on Collaborative Emerging Systems

(COLLABES-2019), which will be held in conjunction with the 33rd International
The emergence of large-scale distributed computing paradigms, such as cloud
computing and mobile computing systems, has opened up many opportunities for
collaborative emerging systems (CES). Collaboration services are at the core of any
information system, aiming to support collaboration among individuals,
inter-teams, and intra-teams. While most CES have focused on using groupware
systems (shared workspaces, videoconferencing, etc.) to support sharing and col-
laboration within one domain, the trend is to develop CES capable to support CES
at large-scale integration. The objective of CES is thus to support efficient col-
laboration at large scale by addressing heterogeneity of resources, multi-domain
nature, availability of resources, and system reliability.
The aim of COLLABES workshop is to gather innovative academic research
results and experiences and best practices related to all aspects of collaborative
emerging systems. I thank all of the authors and Reviewers for their support to the
workshop, and we look forward to their contributions in forthcoming editions of the
workshop.
Finally, I would like to thank the Local Arrangement Team for the support and
good local arrangement for the conference. I hope that workshop participants will
have an enjoyable experience in Matsue, Japan.
Leonard Barolli
COLLABES-2019 Organizer
xxxi
xxxii Welcome Message from COLLABES-2019 International Workshop Organizer
COLLABES-2019 Organizing Committee

Workshop Organizer
Fatos Xhafa Technical University of Catalonia, Spain

Keita Matsuo Fukuoka Institute of Technology, Japan
Serguei Dobrinevski Hypersoft, Germany
Ciprian Dobre Politehnica University of Bucharest, Romania
Natalia Kryvinska University of Vienna, Austria
Hiroaki Nishino University of Oita, Japan
Markus Aleksy ABB AG Corporate Research Center, Germany
Marek Ogiela AGH University of Science and Technology, Poland
Lidia Ogiela AGH University of Science and Technology, Poland
Welcome Message from INTRICATE-SEC-2019
Workshop Chair
Welcome to the 6th International Workshop on Security Intricacies in

Cyber-Physical Systems (INTRICATE-SEC-2019). The INTRICATE-SEC-2019
will be held in conjunction with the 33rd International Conference on Advanced
Information Networking and Applications (AINA-2019) at Kunibiki Messe,
Matsue, Japan, from March 27–29, 2019.
The INTRICATE-SEC workshop aims to bring together security practitioners
and researchers to share ideas and research results in the emerging area of infor-
mation security. It also provides a platform to discuss applications of security. The
papers included in the proceedings cover novel theories, designs, and applications
of information security to network applications as well as in advanced distributed
system environments.
Organizing an international workshop requires the support and help of many
people. I would like to thank all the authors for the paper submissions. I appreciate
the support of the Program Committee members and Reviewers who carried out the
most difficult work of carefully evaluating the submitted papers.
The special acknowledgments go to Prof. Makoto Takizawa and Prof. Leonard
Barolli, the Steering Committee Chairs of AINA International Conference, for their
strong encouragement and guidance in organizing the workshop. I would also like
to thank General Co-chairs, Program Co-chairs, and Workshops Chairs for their
support and help in preparing the technical program.
Finally, I would like to thank the Local Arrangement Team for the support and
good local arrangement for the conference. I do hope that you will have a great time
in Matsue, Japan.
Anne Kayem
INTRICATE-SEC-2019 Workshop Chair
xxxiii
xxxiv Welcome Message from INTRICATE-SEC-2019 Workshop Chair
INTRICATE-SEC-2019 Organizing Committee

Workshop Chair
Anne Kayem HPI, Germany
Stephen Wolthusen Royal Holloway, University of London,

UK and Norwegian University of Science
and Technology, Norway
Christoph Meinel HPI, Germany
Stephen Marsh University of Ontario Institute of Technology, Canada
Cristina Alcaraz University of Malaga, Spain
Zeyar Aung Masdar Institute of Science and Technology, UAE
Dieter Hutter DFKI GmbH, Germany
Zekeriya Erkin Delft University of Technology, The Netherlands
Ingo Stengel University of Applied Sciences Karlsruhe, Germany
George Yee Carleton University, Canada
Welcome Message from E3WSN-2019
Welcome to the 5th on Engineering Energy Efficient Internet Worked Smart seNsor
(E3WSN-2019), which is held in conjunction with the 33rd International
This workshop aims at bringing together academics and practitioners from dif-
ferent areas (such as software engineering, system engineering, energy efficiency,
and embedded systems) to promote design, validation, and implementation of
energy-efficient inter-networked smart sensors. The papers included in the pro-
ceedings cover novel languages, designs, simulation, and applications of
inter-networked smart sensors.
E3WSN-2019 contains high-quality research papers submitted by researchers
from all over the world. Each submitted paper was peer-reviewed by Reviewers
who are experts in the subject area of the paper. Based on the review results, the
Program Committee accepted good-quality papers.
For organizing an international workshop, the support and help of many people
are needed. First, we would like to thank all authors for submitting their papers. We
also appreciate the support from Program Committee members and Reviewers who
carried out the most difficult work of carefully evaluating the submitted papers.
The E3WSN organization team would like to give its special thanks to the
Steering Committee Chairs, General Co-chairs, Program Co-chairs, and Workshop
Organizers for their support to organize the workshop.
Leonardo Mostarda
Jonathan Loo
E3WSN-2019 Co-chairs
xxxv
xxxvi Welcome Message from E3WSN-2019 International Workshop Co-chairs
E3WSN-2019 Organizing Committee

Workshop Co-chairs
Leonardo Mostarda Camerino University, Italy

Jonathan Loo University of West London, UK
Dajana Cassioli L’Aquila University, Italy

Mauro Caporuscio Linnaeus University, Vaxjo, Sweden
Changyu Dong New Castle University, UK
Naranker Dulay Imperial College London, UK
Enver Ever Middle East Technical University,
Northern Cyprus Campus
Sandu Florin Transilvania University of Brasov, Romania
Orhan Gemikonakli Middlesex University, UK
Altan Kocyigit Middle East Technical University, Turkey
Ivano Malavolta Vrije Universiteit Amsterdam, The Netherlands
Diletta Cacciagrano Camerino University, Italy
Henry Muccini L’Aquila University, Italy
Alfredo Navarra University of Perugia, Italy
Franco Raimondi Middlesex University London, UK
Purav Shah Middlesex University London, UK
Cornel Stanca Transilvania University of Brasov, Romania
Rosario Culmone Camerino University, Italy
Ramona Trestian Middlesex University, UK
Adnan Yazici Middle East Technical University, Turkey
Krishna Doddapaneni Altiux Innovations, Mountain View, California, USA
Fadi Al-Turjman Middle East Technical University,
Northern Cyprus Campus
Welcome Message from WITIN-2019
International Workshop Organizers
Information and networking technologies have been a major driving force for the
advances of human society for decades. This 4th edition of International Workshop
on Innovative Technologies in Informatics and Networking (WITIN-2019) aims to
attract innovative attempts in the related fields. This workshop will provide a
platform for researchers and practitioners to meet and exchange their ideas.
WITIN-2019 will be held in conjunction with the 33rd International Conference
on Advanced Information Networking and Applications (AINA-2019) at Kunibiki
Messe, Matsue, Japan, from March 27–29, 2019. We accepted very good-quality
papers for WITIN-2019 after a rigorous selection process. We hope all of you will
enjoy the workshop and find it a great opportunity to discuss research works.
Many people contributed to the CFP and paper review of WITIN-2019. We wish
to thank the Program Committee members for their excellent effort in screening and
selecting the papers. We would like to express our gratitude to Prof. Makoto
Takizawa and Prof. Leonard Barolli, the Steering Committee Co-chairs of AINA
International Conference, for their encouragement and guidance. We are very
thankful to General Co-chairs, PC Co-chairs, and Workshop Co-chairs of
AINA-2019, for their great organization of this conference. Last but not least, we
would like to thank and congratulate all the authors for their support to the
workshop.
Hui-Huang Hsu
Yi-Jen Su
WITIN-2019 Workshop Co-chairs
xxxvii
xxxviii Welcome Message from WITIN-2019 International Workshop Organizers
WITIN-2019 Organizing Committee
Workshop Co-chairs

Yi-Jen Su Shu-Te University, Taiwan
Chuan-Yu Chang National Yunlin University of Science

and Technology, Taiwan
Chuan-Wang Chang Kun Shan University, Taiwan
Chao-Ho Chen National Kaohsiung University of Science
Chia-Yen Chen The University of Auckland, New Zealand
Ding-Horng Chen Southern Taiwan University of Science
Ying-Nong Chen National Central University, Taiwan
Wu-Chih Hu National Penghu University of Science
Deng-Yuan Huang Da-Yeh University, Taiwan
Chipan Hwang National Changhua University of Education, Taiwan
Ji-Han Jiang National Formosa University, Taiwan
Chien-Chuan Ko National Chiayi University, Taiwan
Lung-Jen Wang National Pingtung University, Taiwan
Chien-Chung Wu Southern Taiwan University of Science
Welcome Message from BOSON-2019
International Workshop Chair
Welcome to the 4th International Workshop on Big Data Processing in Online

Social Network (BOSON-2019), which is held in conjunction with the 33rd
(AINA-2019), a at Kunibiki Messe, Matsue, Japan, from March 27–29, 2019.
The purpose of BOSON workshop is to bring together scientists, engineers, and
students for sharing experiences, ideas, and research results about big data pro-
cessing and online social networks.
This workshop provides an international forum for researchers and participants.
It aims to share and exchange experiences, discuss challenges, and present original
ideas in all aspects related to the big data processing and online social network
applications design and development. In particular, we have encouraged innovative
contributions about: big data processing, online social network, distributed com-
puting applications, ubiquitous positioning methods, topologies for distributed
computing, modeling and simulation of big data, modeling and simulation of dis-
tributed systems, distributed knowledge management, systems and algorithms for
social search, collaborative platforms, distributed computing for smart cities, dis-
tributed computing for e-health, smart grid modeling; quality evaluation of dis-
tributed services, semantic technologies, modeling social networks and behavior,
management of social network data, information propagation in social networks,
data mining and machine learning in social systems, sentiment analysis on OSNs,
privacy and security in social systems, trust and reputations in social systems.
For organizing an international workshop, the support and help of many people
are needed. First, I would like to thank the Organizing Committee of AINA-2019
International Conference for giving us the opportunity to organize the
workshop. Second, I would like to thank our Program Committee Members and
Reviewers who carried out the most difficult work of carefully evaluating the
submitted papers. I also would like to thank all authors for submitting their research
works and for their participation. Moreover, I would like to thank the Local
Arrangement Chairs of AINA-2019 conference.
I would like to give my special thanks to Prof. Leonard Barolli and Prof. Makoto
Takizawa, the Steering Committee Chairs of AINA, for their strong encouragement
xxxix
xl Welcome Message from BOSON-2019 International Workshop Chair
and guidance to organize the workshop. I would like to thank AINA-2019 General
Co-chairs, PC Co-chairs, and Workshop Co-chairs for their support and help to
prepare the technical program.
I hope you will enjoy BOSON workshop and AINA International Conference,
find this a productive opportunity for sharing experiences, ideas, and research
results with many researchers, and have a great time in Matsue, Japan.
Flora Amato
BOSON-2019 Workshop Chair
Welcome Message from BOSON-2019 International Workshop Chair xli
BOSON-2019 Organizing Committee

Workshop Chair
Flora Amato University of Naples “Federico II,” Italy
Workshop Program Committee Co-chairs
Francesco Moscato University of Campania “Luigi Vanvitelli,” Italy

Vincenzo Moscato University of Naples “Federico II,” Italy
Antonino Mazzeo University of Naples “Federico II,” Italy

Nicola Mazzocca University of Naples “Federico II,” Italy
Carlo Sansone University of Naples “Federico II,” Italy
Beniamino di Martino University of Campania “Luigi Vanvitelli,” Italy
Antonio Picariello University of Naples “Federico II,” Italy
Valeria Vittorini University of Naples “Federico II,” Italy
Anna Rita Fasolino University of Naples “Federico II,” Italy
Umberto Villano University of Sannio, Italy
Kami Makki Lamar University, Beaumont (Texas), USA
Valentina Casola University of Naples “Federico II,” Italy
Porfirio Tramontana University of Naples “Federico II,” Italy
Aniello Castiglione University of Naples “Federico II” and University
of Salerno, Italy
Massimo Esposito Institute for High Performance Computing
and Networking (ICAR), Italy
Luigi Gallo Institute for High Performance Computing
Aniello Minutolo Institute for High Performance Computing
Angelo Esposito Institute for High Performance Computing
Giovanni Cozzolino University of Naples “Federico II,” Italy
Giovanna Sannino Institute for High Performance Computing
and Networking (ICAR)
Fiammetta Marulli Institute for High Performance Computing
and Networking (ICAR)
Marco Pota Institute for High Performance Computing
Giuseppe Caggianese Institute of High Performance Computing
xlii Welcome Message from BOSON-2019 International Workshop Chair
Francesco Mercaldo Consiglio Nazionale delle Ricerche, Italy

Alessandro Maisto University of Salerno, Italy
Serena Pelosi University of Salerno, Italy
Raffaele Guarasci University of Salerno, Italy
Welcome Message from IOEMLA-2019
International Workshop Co-chair
Welcome to the 2nd International Workshop on Internet of Everything and Machine

Learning Applications (IOEMLA-2019), which will be held in conjunction with the
33rd International Conference on Advanced Information Networking and
2019.
The workshop is intended to facilitate exchange of ideas between researchers
and presents the latest developments and research in the areas of IoE and machine
learning in cross-disciplinary domains such as data science, IoT, industrial IoE, and
computer and intelligent networks.
This international workshop collected research papers written by some estab-
lished international researchers. These papers went through multiple review cycles
and were handpicked based on their quality, clarity, and relevance to the theme of
this workshop.
Many people have kindly helped us to prepare and organize the IOEMLA-2019
workshop. Firstly, I would like to thank the authors who submitted high-quality
papers. I also would like to acknowledge the valuable comments of the Reviewers,
which have enabled us to select these papers out of the so many submissions we
received.
I would like to give our special thanks to Prof. Leonard Barolli and Prof. Makoto
Takizawa, as Steering Committee Co-chairs of AINA International Conference, for
their strong encouragement and guidance to organize the workshop. I also would
like to thank AINA-2019 General Co-chairs, PC Co-chairs, and Workshop
Co-chairs for their continuous support throughout the entire process of the
IOEMLA-2019 workshop. I do hope that you will have a wonderful time in Matsue,
Japan.
Omid Ameri Sianaki

IOEMLA-2019 Workshop Co-chair
xliii
xliv Welcome Message from IOEMLA-2019 International Workshop Co-chair
IOEMLA-2019 Organizing Committee

Workshop Co-chair
Omid Ameri Sianaki Victoria University, Australia
Elizabeth Chang UNSW Canberra, Australia

Vidyasagar Potdar Curtin University, Australia
Mahmoud El Khodr CQUniversity, Australia
Ahmed Dawoud Victoria University, Australia
Nedal Ababneh Victoria University, Australia
Khaled Kourouche Victoria University, Australia
Atif Ali Victoria University, Australia
Shafquat Hussain Victoria University, Australia
Mehregan Mahdavi Victoria University, Australia
Morteza Saberi UNSW Canberra, Australia
Thomas Houghton Curtin University, Australia
Yuan Miao Victoria University, Australia
Jakub Szajman Victoria University, Australia
Zoohan Gani Victoria University, Australia
Gitesh Raikundalia Victoria University, Australia
Javid Taheri Karlstad University, Sweden
Artemis Gharagozlu Victoria University, Australia
Khandakar Ahmed Victoria University, Australia
Farshid Hajati Victoria University, Australia
Azadeh Rajabian Victoria University, Australia
Tabesh
Pantea Aria La Trobe University, Australia
Ashkan Yusefi UC Berkeley, USA
Ali Anaissi The University of Sydney, Australia
Mohammadreza The University of Sydney, Australia
Hoseinfarahabady
Israel Casas Lopez Victoria University, Australia
Belal Alsinglawi Western Sydney University, Australia
Welcome Message from M2EC-2019 Workshop
Organizers
Welcome to the 1st International Workshop on Multi-Clouds and Mobile Edge

Computing (M2EC-2019), which will be held in conjunction with the 33rd
(AINA-2019) at Kunibiki Messe, Matsue, Japan, from March 27–29, 2019.
Today, a large number of enterprises and individuals rely on services offered by
clouds to meet their computational and storage demands. However, in general, no
single cloud provider is able to provide all the features a user may need in a
cost-efficient way, while satisfying the user’s security and performance require-
ments. Even the most dominant cloud providers have limited geographical pres-
ence. Cloud federation enables end users to integrate segregated resources from
different cloud providers. The use of multi-clouds offers more freedom to the cloud
users and increases the granularity of choices in the application deployment.
Growing interest in multi-clouds pushes the need to investigate a large number of
under-explored research topics, ranging from multi-cloud resource provisioning,
application deployments, automated configurations, federated networking, adapta-
tions, security and privacy. Moreover, with the growing need of real-time data
analytics and critical event handling by many modern applications, such as in the
Internet of Things (IoT), it is evident that the centralized computation and storage
model offered by cloud computing are not suitable for such applications, due to
high end-to-end latencies. Mobile edge computing (MEC) enables a computing and
storage infrastructure provisioned closely to the end users at the edge of a cellular
network. Combining MEC in multi-cloud infrastructures can help to combat latency
challenges imposed by the cloudcentric architectures.
The intent of this workshop is to bring together people from research and
industry, in order to provide a discussion forum for state-of-the-art topics related to
cloud, multi-cloud, and mobile edge computing technology, networks, and
applications.
Many people have kindly helped us to prepare and organize the M2EC-2019
workshop. First of all, we would like to thank AINA-2019 Organization Committee
for their support, guidance, and help for making the workshop. We would like to
xlv
xlvi Welcome Message from M2EC-2019 Workshop Organizers
express our special thanks to M2EC-2019 Program Committee members and

Reviewers.
Dimitris Apostolou
Thomas Dreibholz
Feroz Zahid
M2EC-2019 Workshop General Co-chairs
Welcome Message from M2EC-2019 Workshop Organizers xlvii
M2EC-2019 Organizing Committee

General Co-chairs
Dimitris Apostolou Institute of Communication and Computer Systems,

Greece
Thomas Dreibholz Simula@OsloMet – Simula Metropolitan Centre
for Digital Engineering, Norway
Feroz Zahid Simula Research Laboratory, Norway
Publicity Chairs
Xing Zhou Hainan University, China

Paweł Skrzypek 7bulls.com, Poland
Xuejun Cai Ericsson, Sweden

Jörg Domaschka Ulm University, Germany
Ahmed Elmokashfi Simula@OsloMet – Simula Metropolitan Centre
Ernst Gunnar Gran NTNU Gjøvik, Norway
Ralph Holz University of Sydney, Australia
Geir Horn University of Oslo, Norway
Quentin Jacquemart CRNS, France
Kyriakos Kritikos FORTH, Institute of Computer Science, Greece
Xiaoyi Lu The Ohio State University, USA
Sabita Maharjan Simula@OsloMet – Simula Metropolitan Centre
Somnath Mazumdar Simula Research Laboratory, Norway
Omer Rana Cardiff University, UK
Samuel Graphcore, Norway
Rodrigo-Mocholí
Jesús Universidad de Castilla-La Mancha, Spain
Escudero-Sahuquillo
Paresh Saxena BITS Pilani University, India
Jawwad Shamsi FAST NUCES, Pakistan
Tor Skeie Fabriscale Technologies, Norway
Vlado Stankovski University of Ljubljana, Slovenia
Salman Taherizadeh Jozef Stefan Institute, Slovenia
Amir Taherkordi University of Oslo, Norway
Yiannis Verginadis Institute of Communication and Computer Systems,
Greece
Øyvind Ytrehus University of Bergen, Norway
Welcome Message from AIMAL-2019 Workshop
Chair
Welcome to the 1st International Workshop on Artificial Intelligence and Machine

Learning (AIMAL-2019), which will be held in conjunction with the 33rd
(AINA-2019) at Kunibiki Messe, Matsue, Japan, from March 27–29, 2019.
The scope of the AIMAL is to provide a platform for sharing the work from the
domains of smart grid (SG), cloud computing (CC), and big data (BD). This
platform enables the collaboration of the researchers around the globe via sharing
their work in the aforementioned domains. SG acts as the backbone platform for the
integration and development of new tools and technologies using the CC and BD
methodologies. It covers the supply-side and demand-side electricity management
through bidirectional communication mechanisms. Supply side relies on the man-
agement of electricity generation, transmission, and distribution, whereas demand
side focuses on the consumers’ energy consumption management. SG also
describes the demand response techniques in order to facilitate the consumers in
terms of minimizing their electricity bills and energy consumption by maintaining
the user preferences using the CC and BD technologies.
The goal of this workshop is to collect novel academic research work including
its simulation and best experimental results regarding all perspectives of the SG,
CC, and BD applications.
Many people supported and encouraged us to prepare and organize the
AIMAL-2019 workshop. We would like to thank AINA-2019 Organization
Committee for their guidance and help. We would like to express our special thanks
to AIMAL-2019 Program Committee members and Reviewers.
Nadeem Javaid
AIMAL-2019 Workshop Chair
xlix
l Welcome Message from AIMAL-2019 Workshop Chair
AIMAL-2019 Organizing Committee

Workshop Chair
Kamran Munir University of the West England, UK

Safdar Hussain Bouk DGIST, Korea
Muhammad Imran King Saud University, Saudi Arabia
Syed Hassan Ahmed Georgia Southern University, USA
Hina Nasir Air University Islamabad, Pakistan
Sakeena Javaid COMSATS University Islamabad, Pakistan
Rasool Bakhsh COMSATS University Islamabad, Pakistan
Asif Khan COMSATS University Islamabad, Pakistan
Adia Khalid COMSATS University Islamabad, Pakistan
Sana Mujeeb COMSATS University Islamabad, Pakistan
AINA-2019 Keynote Talks
Utilizing Wireless Communication to Enable New
Types of Industrial Applications
Markus Aleksy
ABB AG Corporate Research Center, Ladenburg, Germany
Abstract. The recent progress in the development of wireless communication

technologies enables new types of industrial applications. Traditional industrial
systems usually rely on wired communication technologies to connect sensors and
actuators. However, these type of communication needs to be replaced by wireless
technologies in future to address new developments, such as mixed reality appli-
cations, automated guided vehicles, moving robots and drones or achieving higher
flexibility required by increasing demand for highly customized products and
adaptable production facilities. In this talk, we will address and discuss represen-
tative use cases and concepts focusing on the usage of wireless technologies in an
industrial setting. Moreover, we will present the related challenges and require-
ments of communication networks in such environments and discuss the applica-
bility of 5th generation wireless communication systems.
liii
Lévy Walk on Graphs: Message Dissemination
and Uninformed Search
Naohiro Hayashibara
Kyoto Sangyo University, Kyoto, Japan
Abstract. Random walks play an important role in computer science, spreading a

wide range of topics in theory and practice, including networking, distributed
systems, and optimization. Particularly, Lévy walk, a family of random walks, has
attracted attention in recent years in the context of mobile ad-hoc networks,
delay-tolerant networks, opportunistic communication, and global optimization
problems. It is also used as a model of various things not only in informatics but
also in biology and environmental science. Lévy walk is a mathematical fractal
which is characterized by long segments followed by shorter hops in random
directions. More precisely, the step distance obeys the power law distribution. The
pattern has been found by Paul Lévy, but the similar pattern has also been evolved
as a naturally selected strategy that gives animals and insects an edge in the search
for sparse targets to survive. In fact, this movement pattern has been observed in the
molecular machinery operating in cells, bacteria, the behavior of honeybees,
mussels, mud snails, wandering albatross and shearwaters. In the area of computer
science, it is most likely to be used as the mobility model in mobile ad-hoc net-
works because of the statistical similarity of human mobility. Most of the research
work, however, assumes a continuous plane and hardly any results on graphs are
available. The goal of this keynote is introducing Lévy walk and its variants and
presenting the challenge on Geometric graphs, especially Unit disk graphs,
regarding message dissemination and uninformed search by using Lévy walk. The
results on both message dissemination and uninformed search show that Lévy walk
is quite efficient compared to random walks because of its ballistic trajectory. They
also clarify the relationship between the efficiency of message dissemination and
uninformed search, and the average degree/the diameter of the graphs.
lv
Contents
The 15th International Symposium on Frontiers of Information

Systems and Network Applications (FINA-2019)
Adaptive Waveform Design of Polarimetric Radar for Extended
Targets in Signal-Dependent Clutter . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
Mengxin Yuan, Xu Cheng, Jing Zhang, and Xiaodong Tan
CRAWL: A Trace Routing Algorithm Based on Hybrid
Two-Layer Topology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
Li-ming Zheng, Wen-feng Long, Yu-Jia Liu, and Wei-dong Sun
Design and Implementation of IoT Based Class Attendance
Monitoring System Using Computer Vision and Embedded
Linux Platform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
Hasan Salman, Md Nasir Uddin, Samuel Acheampong, and He Xu
An Indoor 3D Positioning Technology Based on NB-IoT . . . . . . . . . . . . 35
Donghui Xue, He Xu, and Peng Li
Consideration of Implementation Method for the Cloud Type
Virtual Policy Based Network Management Scheme
for the Specific Domain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
Kazuya Odagiri, Shogo Shimizu, Naohiro Ishii, and Makoto Takizawa
A Model of Virtual Machines to Support Storage Processes . . . . . . . . . 57
Naohiro Ogura, Dilawaer Duolikun, Tomoya Enokido,
and Makoto Takizawa
Energy Efficient Scheduling of Smart Home . . . . . . . . . . . . . . . . . . . . . 67
Sajjad Khan, Zahoor Ali Khan, Nadeem Javaid,
Sahibzada Muhammad Shuja, Muhammad Abdullah, and Annas Chand
lvii
lviii Contents
Minimizing Daily Cost and Maximizing User Comfort Using

a New Metaheuristic Technique . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
Raza Abid Abbasi, Nadeem Javaid, Sajjad Khan, Shujat ur Rehman,
Amanullah, Rana Muhammad Asif, and Waleed Ahmad
Efficiency Analysis of TFHE Fully Homomorphic Encryption
Software Library Based on GPU . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
Hai-bin Yang, Wu-jun Yao, Wen-chao Liu, and Bin Wei
NADEEM: A Novel Reliable Data Delivery Routing Protocol
for Underwater WSNs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
Nadeem Javaid
An Architecture for Distributed Ledger-Based M2M Auditing
for Electric Autonomous Vehicles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
Dragos Strugar, Rasheed Hussain, Manuel Mazzara, Victor Rivera,
Ilya Afanasyev, and JooYoung Lee
Is QUIC Quicker Than TCP? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
Saurabh, Satya Prakash, Pranav Kumar Singh, Sunit Kumar Nandi,
and Sukumar Nandi
PRAN: A Provenance Based Model and Prototype to Strengthen
Authentication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
Rajidi Satish Chandra Reddy and Srinivas Reddy Gopu
The Deepest Vertex First Reboot: Rebooting Network Edge
Switches in a Campus Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
Motoyuki Ohmori, Satoshi Fujio, and Masayuki Higashino
Analyzing and Recognizing Pedestrian Motion Using 3D Sensor
Network and Machine Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
Ningping Sun, Toru Tsuruoka, Shunsuke Murakami,
and Takuma Sakamoto
Novel Interestingness Measures for Mining Significant Association
Rules from Imbalanced Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
Safa Abdellatif, Mohamed Ali Ben Hassine, and Sadok Ben Yahia
Extended Scheme Mediation Integration Model for Information
Systems Project Proposal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
William Chaves de Souza Carvalho, Pedro Frosi Rosa,
and Flávio de Oliveira Silva
The 15th International Workshop on Heterogeneous Wireless

Networks (HWISE-2019)
An Analysis on Recent Mobile Application Trend in Bangladesh . . . . . 195
Md Anik Hasan, Nazifa Tasneem, Sumaiya Binte Akther, Koushiq Das,
and Ashik Mostafa Alvi
Contents lix
Performance Evaluation of Routing Protocols in DTNs Considering

Different Mobility Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
Evjola Spaho, Klodian Dhoska, Kevin Bylykbashi, Leonard Barolli,
Vladi Kolici, and Makoto Takizawa
A Distance-Based Advertisement-Delivery Method
for Vehicular DTN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215
Shogo Nakasaki, Yu Yoshino, Makoto Ikeda, and Leonard Barolli
Numerical Simulation of Glide Slope Signal Interferences
by Irregular Ground . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 224
Junichi Honda, Hirohisa Tajima, and Hisashi Yokoyama
Effect of Client Priority in the Performance of a Fuzzy-Based
WLAN Triage System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 234
Kosuke Ozera, Yi Liu, Leonard Barolli, and Makoto Takizawa
Energy Saving in HetNet Network Using eNB Parameters Tuning . . . . 244
Narjes Lassoued, Noureddine Boujnah, and Ridha Bouallegue
Robust Road Lane Detection for High Speed Driving
of Autonomous Vehicles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 256
Hyunhee Park
Study of Beam Power Control of Ka-Band Multi-beam
Broadcasting Satellite Using Meteorological Data . . . . . . . . . . . . . . . . . 266
Takumi Iwamoto and Kiyotaka Fujisaki
Numerical Analysis of Optical Duplexer Composed of Dispersive
and Nonlinear Dielectric in Two-Dimensional Photonic Crystal
Waveguide with Square Lattice . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 275
Keisuke Haari, Naoki Higashinaka, Xiang Zheng Meng,
and Hiroshi Maeda
Cycle Interference Alignment for the Full Duplex Communication
System Based on User Virtual Splitting Thought . . . . . . . . . . . . . . . . . . 286
Haiying Ren, Di Wu, Man Li, and Tianyi Feng
The 12th International Symposium on Mining and Web (MAW-2019)

Proposal of Web API Architecture for Smart Community:
A Case Study of Japan . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 295
Kanae Matsui
Evaluation Measures for Extended Association Rules Based
on Distributed Representations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 305
Tomonobu Ozaki
lx Contents
Estimation of Emotion Type and Intensity in Japanese Tweets

Using Multi-task Deep Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 314
Kazuki Sato and Tomonobu Ozaki
A Method for Extracting Influential People for the Improvement
of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 324
Hayato Tsukiji and Kosuke Takano
Sentiment Analysis of Arabic and English Tweets . . . . . . . . . . . . . . . . . 334
Mohamed K. Elhadad, Kin Fun Li, and Fayez Gebali
The 12th International Workshop on Bio and Intelligent

Computing (BICom-2019)
A Deep Q-Network with Experience Optimization (DQN-EO)
for Atari’s Space Invaders . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 351
Yan Chen and Elis Kulla
Design of a Deep Q-Network Based Simulation System
for Actuation Decision in Ambient Intelligence . . . . . . . . . . . . . . . . . . . 362
Tetsuya Oda, Chiaki Ueda, Ryo Ozaki, and Kengo Katayama
An Efficient Scheduling of User Appliances Using Multi
Objective Optimization in Smart Grid . . . . . . . . . . . . . . . . . . . . . . . . . . 371
Hafiz Muhammad Faisal, Nadeem Javaid, Umar Qasim, Shujaat Habib,
Zeshan Iqbal, and Hasnain Mubarak
Pro Utility Pro Consumer Comfort Demand Side Management
in Smart Grid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 385
Waleed Ahmad, Nadeem Javaid, Basit Karim, Syed Qasim Jan,
Muhammad Ali, Raza Abid Abbasi, and Sajjad Khan
Efficient Scheduling of Smart Home Appliances for Energy
Management by Cost and PAR Optimization Algorithm
in Smart Grid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 398
Sahibzada Muhammad Shuja, Nadeem Javaid, Sajjad Khan,
Hina Akmal, Murtaza Hanif, Qazi Fazalullah, and Zain Ahmad Khan
Multiple S-Box Correlation Energy Analysis Model Based
on Particle Swarm Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 412
Wu-jun Yao, Hai-bin Yang, Lin Chen, and Bin Wei
Improving Peer Reliability Considering Jitter Parameter:
A Fuzzy-Based System for JXTA-Overlay P2P System . . . . . . . . . . . . . 422
Yi Liu, Makoto Ikeda, Keita Matsuo, Leonard Barolli,
and Makoto Takizawa
ANN Based Intrusion Detection Model . . . . . . . . . . . . . . . . . . . . . . . . . 433
Seunghyun Park and Hyunhee Park
Contents lxi
The 12th International Workshop on Telecommunication Networking,

Implementation of a Transnational Testbed and Web UI System
with Layer3 SDX . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 441
Wun-Yuan Huang, Hui-Lan Lee, Ta-Yuan Chou, Te-Lung Liu, Fei Yeh,
Jim Hao Chen, and Joe Mambretti
The Case Study of Software Build-in Design Based on Quality
Factors and FMEA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 451
Meng-Ling Hsieh, Wei-Tsen Lin, Suhan Yu, Yi-Chi Chen, Jung-Shan Lin,
and Lin-Hui Nung
A Novel Sharing M-Coupons with Lightweight Computations
via Cloud Computing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 459
Iuon-Chang Lin and Hsiao-Chi Chiang
A Smart Roadside Parking System Using Bluetooth Low
Energy Beacons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 471
Hui-Tzu Chen, Pei-Yu Lin, and Chi-Yi Lin
A Light Weight Stream Cypher Mechanism for Visible Light
Communication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 481
Shih-Hao Chang, Ted Huang, and Mei-Lan Chen
The 10th International Workshop on Disaster and Emergency

Information Network Systems (IWDENS-2019)
Delay Tolerant Networks with Static Body Object Detections
by Mobile Sensors for Disaster Information System . . . . . . . . . . . . . . . . 493
Noriki Uchida, Tomoyuki Ishida, and Yoshitaka Shibata
A Basic Study on Emergency Communication System
for Disaster Using LPWA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 501
Hiroaki Yuze and Shinichi Nabeta
Platform System Based on LoRa Mesh Network Technology . . . . . . . . . 510
Goshi Sato, Yoshitaka Shibata, and Noriki Uchida
Construction of a Disaster Response Automatic Extraction
Support System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 516
Tatsuya Ohyanagi, Tomoyuki Ishida, Noriki Uchida,
and Yoshitaka Shibata
Network Performance Evaluation to Realize N-Wavelength V2X
Cognitive Wireless Communication System . . . . . . . . . . . . . . . . . . . . . . 524
Akira Sakuraba, Takanori Ito, and Yoshitaka Shibata
lxii Contents
A IoT Based Disaster Information Platform for Challenged

Network Environment in Snow Countries . . . . . . . . . . . . . . . . . . . . . . . 537
Yoshitaka Shibata, Yoshikazu Arai, Yoshia Saito, and Jun Hakura
Study on Balloon Network Using LoRa Mesh
Communication System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 545
Goshi Sato, Yoshitaka Shibata, and Noriki Uchida
Performance Evaluations on Adaptive Array Antenna of Vehicular
Delay Tolerant Networking for Winter Road Surveillance Systems . . . . 550
Noriki Uchida, Goshi Sato, and Yoshitaka Shibata
The 7th International Workshop on Collaborative Emerging Systems

(COLLABES-2019)
Design of Modern Logistics Management System Based on RFID
and NB-IoT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 561
Jiayi Pang, Leixian Shen, Qingyun Zhang, He Xu, and Peng Li
Evaluation of TBC and OBC Precedent Relations Among Messages
in P2P Type of Topic-Based Publish/Subscribe System . . . . . . . . . . . . . 570
Takumi Saito, Shigenari Nakamura, Dilawaer Duolikun, Tomoya Enokido,
and Makoto Takizawa
Data Analytics for Electricity Load and Price Forecasting
in the Smart Grid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 582
Syeda Aimal, Nadeem Javaid, Amjad Rehman, Nasir Ayub,
Tanzeela Sultana, and Aroosa Tahir
An Efficient CNN and KNN Data Analytics for Electricity
Load Forecasting in the Smart Grid . . . . . . . . . . . . . . . . . . . . . . . . . . . 592
Syeda Aimal, Nadeem Javaid, Tahir Islam, Wazir Zada Khan,
Mohammed Y. Aalsalem, and Hassan Sajjad
A New System for Management of IoT Sensors Considering
Agile-Kanban . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 604
Keita Matsuo, Takeru Kurita, and Leonard Barolli
A Two-Way Flow Model for Fog Computing . . . . . . . . . . . . . . . . . . . . . 612
Yinzhe Guo, Ryuji Oma, Shigenari Nakamura, Dilawaer Duolikun,
Tomoya Enokido, and Makoto Takizawa
The 6th International Workshop on Security Intricacies in

Cyber-Physical Systems and Services (INTRICATE-SEC-2019)
Stochastic Methods to Find Maximum Likelihood for Spam
E-mail Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 623
Seyed M.-H. Mansourbeigi
Contents lxiii
Design and Research of Trusted Acquisition Terminals Based

on Domestic Password . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 633
Yun Feng, Xiaobing Liang, Feng Zhai, and Meng Xu
Trust-Based Security Mechanism for Detecting Clusters of Fake
Users in Social Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 641
Davinder Kaur, Suleyman Uslu, and Arjan Durresi
Machine Learning Based Approach to Detect Wormhole Attack
in VANETs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 651
Pranav Kumar Singh, Rahul Raj Gupta, Sunit Kumar Nandi,
and Sukumar Nandi
Secure Peer-to-Peer Communication Based on Blockchain . . . . . . . . . . . 662
Kahina Khacef and Guy Pujolle
Developing the Analysis Tool of Cyber-Attacks by Using CTI
and Attributes of Organization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 673
Yusuke Kambara, Yoshinori Katayama, Takanori Oikawa,
Kazuyoshi Furukawa, Satoru Torii, and Tetsuya Izu
Proposal of Ad-Hoc Secure Device Pairing Method Using Similarity
Between Marker Movement and Acceleration . . . . . . . . . . . . . . . . . . . . 683
Makoto Nagatomo, Kentaro Aburada, Hisaaki Yamaba, Naonobu Okazaki,
and Mirang Park
Introduction of Fingerspelling for Realizing a User Authentication
Method Using s-EMG . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 693
Hisaaki Yamaba, Shimpei Inotani, Shotaro Usuzaki, Kayoko Takatsuka,
Kentaro Aburada, Tetsuro Katayama, Mirang Park, and Naonobu Okazaki
App-Collusion Detection Using a Two-Stage Classifier . . . . . . . . . . . . . . 702
Md. Faiz Iqbal Faiz, Md. Anwar Hussain, and Ningrinla Marchang
The 5th International Workshop on Engineering Energy Efficient

InternetWorked Smart seNsors (E3WSN-2019)
Comparison of Machine Learning Techniques for Prediction
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 713
Yoney Kirsal Ever, Kamil Dimililer, and Boran Sekeroglu
Tailoring Micro-solar Systems to Heterogeneous Wireless
Sensor Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 724
Stefano Calabrò, Roberto Gagliardi, Fausto Marcantoni, Matteo Micheletti,
Alessandro Pacini, and Andrea Piermarteri
Distributing Energy Consumption in Multi-interface
Series-Parallel Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 734
Alessandro Aloisio, Alfredo Navarra, and Leonardo Mostarda
lxiv Contents
Energy Efficient Light Routing in Utility Network . . . . . . . . . . . . . . . . . 745

Rosario Culmone and Fabio Pagnotta
The 4th International Workshop on Innovative Technologies

in Informatics and Networking (WITIN-2019)
Performance of the 25 Gbps/100 Gbps Fullmesh RoCE Network
Using Mellanox ConnetX-4 Lx Adapter and Ruijie S6500
Ethernet Switch . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 757
Hui Li, Xiaolong Chen, Tao Song, Haiyin Chen, and Hao Chen
The Warning System for Speed Cameras on the Road
by Deep Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 768
Chien-Chung Wu, Yu-Xuan Lin, Deng-Xiang Hu, Chien-Chuan Ko,
and Ji-Han Jiang
Using Feature Selection to Improve Performance of Three-Tier
Intrusion Detection System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 776
Yi-Jen Su, Pei-Yu Huang, Wu-Chih Hu, Hsuan-Yu Lin, Chen-Yu Kao,
Shan-Hsiung Hsieh, and Chun-Li Lin
A CNN-Based Method for Infant Cry Detection and Recognition . . . . . 786
Chuan-Yu Chang and Lung-Yu Tsai
The Sensor Calibration and Growth Parameters Monitoring
for Phalaenopsis Cultivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 793
Ding-Horng Chen, Rong-Show Shen, Tsai-Rong Chang, Pei-Shan Lin,
and Tzu-Ying Wang
The 4th International Workshop on Big Data Processing

in Online Social Network (BOSON-2019)
Automatic Text Preprocessing for Intelligent Dialog Agents . . . . . . . . . 805
Alessandro Maisto, Serena Pelosi, Massimiliano Polito,
and Michele Stingo
Exploiting Figures of Speech in Cultural Heritage Reasoning . . . . . . . . 815
Flora Amato, Walter Balzano, Giovanni Cozzolino,
and Francesco Moscato
Textual Processing in Social Network Analysis . . . . . . . . . . . . . . . . . . . 822
Flora Amato, Walter Balzano, Giovanni Cozzolino, Alessandro de Luca,
and Francesco Moscato
ACOp: An Algorithm Based on Ant Colony Optimization
for Parking Slot Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 833
Walter Balzano and Silvia Stranieri
Contents lxv
A Self-organization Technique in Wireless Sensor Networks

to Address Node Crashes Problem and Guarantee
Network Connectivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 841
Data Dissemination in Vehicular Ad Hoc Network: A Model
to Improve Network Congestion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 851
ManDri: A New Proposal of Manus VR Facility Integration
in Everyday Car Driving . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 860
Walter Balzano, Maurizio Minieri, and Silvia Stranieri
A Smart Compact Traffic Network Vision Based on Wave
Representation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 870
Walter Balzano, Aniello Murano, Loredana Sorrentino,
and Silvia Stranieri
Big Data Analytics for Traceability in Food Supply Chain . . . . . . . . . . 880
Alessandra Amato, Giovanni Cozzolino, and Vincenzo Moscato
The 2nd International Workshop on Internet of Everything

and Machine Learning Applications (IOEMLA-2019)
Social Credibility Incorporating Semantic Analysis and Machine
Learning: A Survey of the State-of-the-Art and Future
Research Directions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 887
Bilal Abu-Salih, Bushra Bremie, Pornpit Wongthongtham, Kevin Duan,
Tomayess Issa, Kit Yan Chan, Mohammad Alhabashneh,
Teshreen Albtoush, Sulaiman Alqahtani, Abdullah Alqahtani,
Muteeb Alahmari, Naser Alshareef, and Abdulaziz Albahlal
Source Codes Classification Using a Modified Instruction
Count Pass . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 897
Omar Darwish, Majdi Maabreh, Ola Karajeh, and Belal Alsinglawi
Predictive Analytics and Deep Learning Techniques in Electronic
Medical Records: Recent Advancements and Future Direction . . . . . . . 907
Belal Alsinglawi and Omar Mubin
Big Data Analytics for Electricity Price Forecast . . . . . . . . . . . . . . . . . . 915
Ashkan Yousefi, Omid Ameri Sianaki, and Tony Jan
Queue Formation Augmented with Particle Swarm Optimisation
to Improve Waiting Time in Airport Security Screening . . . . . . . . . . . . 923
Mohamad Naji, Ahmed Al-Ani, Ali Braytee, Ali Anaissi,
and Paul Kennedy
lxvi Contents
Polar Topographic Derivatives for 3D Face Recognition: Application

to Internet of Things Security . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 936
Farshid Hajati, Ali Cheraghian, Omid Ameri Sianaki, Behnam Zeinali,
and Soheila Gheisari
A Survey on Conversational Agents/Chatbots Classification
and Design Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 946
Shafquat Hussain, Omid Ameri Sianaki, and Nedal Ababneh
Dimensionality Reduction for Network Anomalies Detection:
A Deep Learning Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 957
Ahmed Dawoud, Seyed Shahristani, and Chun Raun
Blockchain: A Research Framework for Data Security and Privacy . . . 966
Farhad Daneshgar, Omid Ameri Sianaki, and Prabhat Guruwacharya
Environmental Monitoring Intelligent System Using Wireless
Nanosensor Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 975
Nedal Ababneh, Omid Ameri Sianaki, and Shafquat Hussain
The 1st International Workshop on Multi-clouds and Mobile

Edge Computing (M2EC-2019)
Situation Detection on the Edge . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 985
Nikos Papageorgiou, Dimitris Apostolou, Yiannis Verginadis,
Andreas Tsagkaropoulos, and Gregoris Mentzas
A Context-Aware Service for Authorizing Multi-cloud Deployments . . . 996
Yiannis Verginadis, Ioannis Patiniotakis, and Gregoris Mentzas
A Real-Time Video Streaming System over
IPv6+MPTCP Technology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1007
Yu Luo, Xing Zhou, Thomas Dreibholz, and Hanbao Kuang
Towards Realistic Simulations of Arbitrary Cross-Cloud Workloads . . . 1020
Nicolay Mohebi and Feroz Zahid
Data Center Clustering for Geographically Distributed
Cloud Deployments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1030
Dipesh Pradhan and Feroz Zahid
Cost Benefits of Multi-cloud Deployment of Dynamic Computational
Intelligence Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1041
Geir Horn, Paweł Skrzypek, Katarzyna Materka, and Tomasz Przeździȩk
An Overview of Multi-cloud Computing . . . . . . . . . . . . . . . . . . . . . . . . 1055
Jiangshui Hong, Thomas Dreibholz, Joseph Adam Schenkel,
and Jiaxi Alessia Hu
Contents lxvii
The 1st Workshop on Artificial Intelligence and Machine Learning

(AIMAL-2019)
Multi-objective Optimal Power Flow Using Improved Multi-objective
Multi-verse Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1071
Muhammad Abdullah, Nadeem Javaid, Annas Chand, Zain Ahmad Khan,
Muhammad Waqas, and Zeeshan Abbas
Enhanced Robustness Strategy for IoT in Smart Cities Based
on Data Driven Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1084
Talha Naeem Qureshi, Nadeem Javaid, Suhail Ashfaq Butt, Wahab Khan,
Sabir Ali Changazi, and Muhammad Mudassar Iqbal
Game-Theoretical Energy Management for Residential User
and Micro Grid for Optimum Sizing of Photo Voltaic Battery
Systems and Energy Prices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1097
Aqdas Naz, Nadeem Javaid, Abdul Basit Majeed Khan,
Muhammad Mudassar Iqbal, Muhammad Aqeel ur Rehman Hashmi,
and Raheel Ahmad Abbasi
Electricity Load Forecasting for Each Day of Week
Using Deep CNN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1107
Sajjad Khan, Nadeem Javaid, Annas Chand, Abdul Basit Majeed Khan,
Fahad Rashid, and Imran Uddin Afridi
Short Term Load Forecasting Using XGBoost . . . . . . . . . . . . . . . . . . . . 1120
Raza Abid Abbasi, Nadeem Javaid, Muhammad Nauman Javid Ghuman,
Zahoor Ali Khan, Shujat Ur Rehman, and Amanullah
Electricity Price Forecasting in Smart Grid: A Novel E-CNN Model . . . 1132
Waleed Ahmad, Nadeem Javaid, Annas Chand, Syed Yousaf Raza Shah,
Umar Yasin, Mudassar Khan, and Aimal Syeda
Electricity Price Prediction by Enhanced Combination of
Autoregression Moving Average and Kernal Extreme
Learing Machine . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1145
Sahibzada Muhammad Shuja, Nadeem Javaid, Sajjad Khan, Umair Sarfraz,
Syed Hamza Ali, Muhammad Taha, and Tahir Mehmood
Prediction of Building Energy Consumption Using Enhance
Convolutional Neural Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1157
Hafiz Muhammad Faisal, Nadeem Javaid, Bakhtawar Sarfraz, Abdul Baqi,
Muhammad Bilal, Inzamam Haider, and Sahibzada Muhammad Shuja
Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1169
The 15th International Symposium on
Frontiers of Information Systems and
Network Applications (FINA-2019)
Adaptive Waveform Design
of Polarimetric Radar for Extended
Targets in Signal-Dependent Clutter
Mengxin Yuan1(B) , Xu Cheng1 , Jing Zhang1 , and Xiaodong Tan2

1
Department of Information and Communication,
Armed Police College of PAP, Chengdu 610213, People’s Republic of China
243453175@qq.com
2
School of Mechanical and Electrical Engineering,
University of Electronic Science and Technology of China,
Chengdu 611731, People’s Republic of China
Abstract. We focus on the joint design of the transmit waveform and

receive filtering structure for polarimetric radar, with considering the
output Signal to Interference plus Noise Ratio (SINR) as the figure of
merit, under both a similarity and an energy constraint on the transmit
signal. We develop an iterative optimization procedure that monotoni-
cally improves the SINR and converges to a stationary point. Each itera-
tion of the algorithm requires the solution of both a convex and a hidden
convex optimization problem. The resulting computational complexity
is linear with the number of iterations and polynomial with the receive
filter length. Experimental results show the effectiveness of the proposed
technique.
1 Introduction
Recently the increased capability of modern radar systems has enabled the use
of very complex and effective signal processing. This has allowed the radar to
successfully adapt the transmit waveform and receive filter to the operating envi-
ronment in a very sophisticated manner. One attractive problem surely consists
in waveform design for extended targets motivated by the high potentials of
modern High-Resolution Radar (HRR) systems. Moreover, polarization diver-
sity can potentially improve the radar performance w.r.t. conventional radar
systems. Therefore, based on the aforementioned considerations, in this paper
we consider the problem of jointly optimizing the transmit waveform and receive
filter of polarimetric radars for the extended target in clutter.
Supported by the Foundation of Science & Technology Department of Sichuan Province

(No. 2015FZ0071), the Foundation of Education Department of Sichuan Province (No.
16ZA0450).
c Springer Nature Switzerland AG 2019
L. Barolli et al. (Eds.): WAINA 2019, AISC 927, pp. 3–13, 2019.
https://doi.org/10.1007/978-3-030-15035-8_1
4 M. Yuan et al.
Several studies concerning waveform design of polarimetric radar have

appeared in the last decades. Precisely, in [4] an algorithm for the design of opti-
mized fully-polarimetric transmit waveform and receive filter for an extended
target in signal-dependent clutter, is devised. The results there show a signifi-
cant performance improvement of polarization optimization against conventional
fully-polarimetric radar systems. However, the iterative optimization technique
developed in [4] is not guaranteed to converge to an optimal solution. To solve
such drawback, an improved algorithm, guaranteeing non-decreasing SINR at
each iteration and converging to a steady-state value, is proposed in [3]. Never-
theless, it is worthy of noting that, both works [3,4] are based on the following
restrictive assumptions: (1) only energy limitations are accounted in the design of
the transmit radar signal; (2) though the clutter is modelled to be polarization-
correlative, it is assumed range-bin independent.
In this paper, we still deal with the problem of waveform design for polarimet-
ric radars. Other than an energy constraint, a similarity constraint is enforced on
the transmitted radar signal. Besides, we model the clutter correlation in both
polarization and range domains. Considering the SINR as the figure of merit,
the original design is reformulated into an equivalent form which paves the way
to the development of a novel iterative optimization procedure monotonically
improving the SINR. Each iteration of the algorithm, whose convergence to a
stationary point is analytically proved, requires the solution of both a convex
and a hidden-convex fractional problems. Finally, several experimental examples
illustrate the effectiveness of the new method1 .
2 System Model
Assuming a monostatic HRR being employed, the overall scattering behavior of
the probed target is completely specified by the TIRM. It is well known that the
Target Impulse Response Matrix (TIRM) is strongly dependent on the TAA [5],
namely the target scattering behavior is sensitive with respect to even a small
change of the TAA θ. Therefore, taking into account TIRM variability w.r.t.
1
We adopt the notation RN , CN and HN respectively for the sets of N -dimensional
vectors of real numbers, the sets of N -dimensional vectors of complex numbers and
N ×N Hermitian matrices. CN ×N denotes N ×N complex matrices. We use boldface
for vectors a (lower case) and matrices A (upper case). The curled inequality symbol
(and its strict form ) is used to denote generalized matrix inequality: for any
A ∈ HN , A 0 means that A is a positive semidefinite matrix (A 0 for positive
definiteness). The notation A(i, j) represents the i-th row and j-th column element
of the matrix A. I and 0 denote, respectively, the identity matrix and the matrix
(or vector) of zero entries, with their dimensions being determined from the context.
The Euclidean norm of the vector x is denoted by ||x || and for the matrix B,
B2 and BF represent its induced 2-norm and Frobenius norm, respectively. The
symbols (·)∗ , (·)T , (·)† , E[·], (·) and ⊗ denote the conjugate, transpose, Hermitian,
expectation, real part and Kronecker product operators, respectively. Finally, for any
optimization problem P, v(P) and Ω(P) denote the corresponding optimal value and
feasible set, respectively.
Adaptive Waveform Design of Polarimetric Radar 5
TAA represents the key point for an effective transmit-receive design2 . As a

consequence, we will explicitly stress the dependence of the TIRM with respect
to the TAA in the following.
Consider a fast-time full-polarization transmit code consisting of N equally
spaced temporal samples as follows
T
s sH,0 , sV,0 , . . . , sH,N −1 , sV,N −1 ∈ C2N , (1)
where the subscripts H and V denote the horizontal and vertical polarization
components, respectively.
Let the support interval of the TIRM (corresponding to the TAA θ) have
length Q and the observations be collected over a discrete interval M , where
M = Q + N − 1 so that the majority of the extended target echoes is received.
For a certain TAA θ and range bin n, the full-polarization target scattering
matrix T n (θ) ∈ C2×2 , being a matrix-valued sample of the TIRM T (θ), is
expressed as
T (θ) THV,n (θ)
T n (θ) HH,n , (2)
TV H,n (θ) TV V,n (θ)
where θ ∈ [0, 2π), n ∈ {0, 1, . . . , Q − 1}, and TXY,n denotes the scattering coeffi-
cient experienced by transmitted “X” polarization component received on “Y ”
polarization at n-th range bin. Similarly, the clutter scattering matrix associated
to the n-th range bin, being an ensemble of the Clutter Impulse Response Matrix
(CIRM) C , is defined as

CHH,n CHV,n
Cn , n ∈ {−N + 1, . . . , M − 1}. (3)
CV H,n CV V,n
After stacking all the observations into a single vector as

T
r rH,0 , rV,0 , . . . , rH,M −1 , rV,M −1 ∈ C2M ,
the signal model in compact form is given by
r = αT T (θ)s + c + v = αT T (θ)s + C s + v , (4)
where αT is a complex parameter that takes into account radar equation, prop-
agation effects, etc, and v is the 2M -dimensional additive noise
T
v vH,0 , vV,0 , . . . , vH,M −1 , vV,M −1 ∈ C2M . (5)
2
Generally, it is hard to obtain the exact knowledge of target scattering behavior
before the real illumination. However, by employing a cognitive radar theory, it is
reasonable to suppose that the radar system has access to a (potentially dynamic)
database containing a priori electromagnetic reflectivity of the target. Meanwhile,
the TAA can also be considered as a prior known by estimating the target attitude
to the radar.
6 M. Yuan et al.
Q−1 M −1
The matrices T (θ) n=0 J n ⊗ T n (θ) and C n=−N +1 J n ⊗ C n are the
2M × 2N -dimensional TIRM and CIRM, respectively, with J n being the M ×
1 if 1 − 2 = n
N -dimensional shift matrix defined as J n (1 , 2 ) , 1 ∈
0 if 1 − 2 = n
{1, . . . , M }, 2 ∈ {1, . . . , N }. In the following, we will assume that the clutter
and noise are uncorrelated. Let the noise vector v be zero-mean (i.e., E[v ] =
0), and circularly symmetric with covariance matrix Σv = E[v v † ] = σv2 I . Let
CHV,n = CV H,n due to the reciprocity of monostatic scenario. We do not make
any (restrictive) assumption on the multivariate statistical characterization of
c. Differently, for the clutter scatterer in the n-th range bin, assume it to be
zero-mean, i.e., E[C n ] = 0 (viz. E[C ] = 0). Also, we account for the clutter
correlation in both polarization and range domains, namely define
∗
E[Cj,n Ck,n ]
rjk (n, n )

, (6)
E[|Cj,n | ] E[|Ck,n |2 ]
2
where j, k ∈ {HH, HV, V V }, as the normalized correlation parameter between

the clutter scatterers of the “j” channel and the “k” channel at n-th and
n -th range bin, respectively. As to the polarization correlation of the clutter,
we exploit the polarimetric clutter model in [7, Sec. II]. After some algebraic
manipulations, the 2M × 2M clutter covariance matrix Σc (s) is obtained3 .
3 Problem Formulation
After filtering the observation vector r through w [w(0), . . . , w(2M − 1)]T ∈
C2M , the resulting system output can be written as
y = w † r = αT w † T (θ)s + w † C s + w † v , (7)
with the corresponding SINR
|αT |2 |w † T (θ)s|2
SINRθ (s, w ) , (8)
E[|w † C s|2 ] + E[|w † v |2 ]
which can be equivalently recast as:
|αT |2 |w † T (θ)s|2
SINRθ (s, w ) = . (9)
w † (Σc (s) + σv2 I ) w
We consider the output SINR as the relevant figure of merit. Moreover, to
control some relevant features of the probing signal (such as range-Doppler res-
olution, variations in the signal modulus, and peak sidelobe level and etc.), in
addition to a constraint on the transmitted energy, i.e., ||s||2 = 1, a similar-
ity constraint (||s − s 0 ||2 ≤ γ) is enforced on the transmit sequence where the
3
The explicit form of the clutter covariance matrix is not shown due to reasons of
space.
parameter 0 ≤ γ < 1 rules the size of the trust hypervolume and s 0 is a prefixed
code (||s 0 ||2 = 1). Thus, the optimization problem for the joint design of trans-
mit code and receive filter of full-polarization radar, under the aforementioned
energy and similarity constraints, can be formulated as
⎧
⎪ |w † T (θ)s|2
⎨ max
s,w w † Σ (s)w +σ 2 w † w
c v
P s.t. s 2
=1 , (10)
⎪
⎩
s − s 0 ≤ γ
2
in which the term |αT |2 is not included since it does not affect the optimization
problem being investigated.
4 Transmit Code and Receive Filter Design

Aimed at finding a good solution for problem P, we develop an optimization
procedure which iteratively optimizes
|w † T (θ)s|2
g(s, w) max . (11)
s,w w † Σc (s)w + σv2 w † w
In an “alternating” fashion, i.e. by keeping s fixed and maximizing w(m) and

vice versa. To do this, we need the optimal solutions s (m) and w(m) to the
(m) (m)
optimization problems Ps and Pw respectively, defined as:
⎧
⎪ |w (m−1)† T (θ)s|2
⎨ max w (m−1)† Σ c (s)w (m−1) +σ2 w (m−1)† w (m−1)
s v
Ps (m)
s.t. s2 = 1 , (12)
⎪
⎩
s − s 0 2 ≤ γ
and
|w † T (θ)s (m) |2
Pw (m) max w † Σ c (s (m) )w +σv2 w † w
. (13)
w
Literatures have shown that the proposed procedure monotonically converges to

the limit value associated with the employed starting point and that any cluster
point is a stationary point [2,6].
4.1 Receive Filter Optimization: Solution to Problem Pw (m )

As to Pw (m) , the following lemma paves the way to the design of the optimal
receive filter structure.
(m)
Lemma 1. An optimal solution to Pw (m) , is equivalent to solve P̄w , defined
as:
min w† Σc s(m) w + σv2 w† w
P̄w
(m)
w , (14)
s.t. (w† T (θ) s(m) ) = 1,
i.e., given optimal solutions w to P̄w , w is an optimal solution to Pw (m) .

(m)
8 M. Yuan et al.
Proof. The proof is similar to that found in [2] and thus it is omitted for the
sake of brevity.
(m)
Then, P̄w is the well-known minimum-variance distorsionless response
(MVDR) [8] problem and its solution is the Capon filter explicitly as
−1
Σc s (m) + σv2 s (m)2 I T (θ) s (m)
w (m) = −1/2 2 . (15)

Σc s (m)
+ σv s I
2 (m) 2
T (θ) s (m)

4.2 Radar Code Optimization: Solution to Problem Ps (m )
Before describing the solution of Ps (m) , we exhibit Lemma 2 as follow:

Lemma 2. The objective function of P can be recast as:
|w† T(θ)s|2
max (16)
s,w s† Σ̄c (w) + σv2 w† wI s
where Σ̄c (w) represents the filter-dependent clutter covariance matrix and its
explicit expression is omitted for the sake of brevity.
Proof. The proof is similar to that found in [10] and thus it is omitted for the
sake of brevity.
Using Lemma 2, Ps (m) is equivalent to P2 :

⎧
⎪ |w (m−1)† T (θ)s|2
⎨ maxs †
s (Σ̄ c (w (m−1) )+σv2 w (m−1)† w (m−1) I )s
P2 s.t. s = 1
2 , (17)
⎪
⎩
s − s 0 ≤ γ
2
This is a fractional quadratic problem and to solve it, we follow the guidelines
in [2]. Introducing the parameters
†
W = w (m−1)† T (θ) w (m−1)† T (θ) , (18)
and
M = Σ̄c∗ (w (m−1) ) + σv2 w (m−1)† w (m−1) I , (19)
we can obtain the homogenized version of P2 as

⎧ ⎛⎡ ⎤⎡ ⎤⎞
⎪
⎪ W 0⎦⎣ss † st∗ ⎦⎠
⎪
⎪ tr ⎝⎣
⎪
⎪ 00 s † t |t|2
⎪
⎪ max ⎛⎡ ⎤⎡ ⎤⎞
⎪
⎪ s,t M 0 ss †
st ∗
⎪
⎪ tr ⎝⎣ ⎦⎣ ⎦⎠
⎪
⎪ 00 s † t |t|2
⎪
⎪ † ∗
⎪
⎨ I − s0 ss st
P2

s.t. tr † ≤0 ,
⎪ −s s
0 0 2
− δ s † t |t|2 (20)
⎪
⎪ I 0 ss † st∗
⎪
⎪
⎪
⎪ tr † =1
0 0 s †t |t| ∗
2
⎪
⎪
⎪
⎪ 0 0 ss st
⎪
⎪
⎪
⎪ tr
0 1 s † t |t|2
=1
⎪
⎪
⎩
s ∈ C ,t ∈ C
2N

Problem P2 and P2 are clearly equivalent.

Moreover, by dropping the rank-one constraint of P2 , we obtain the semidef-
inite programming (SDP) as problem P3 :
⎧
⎪
⎪ tr (Q 1 S )
⎪
⎪ max
⎪
⎪ S tr (Q 0 S )
⎨
s.t. tr (Q 1 S ) ≤ 0
P3 , (21)
⎪
⎪ tr (Q 2 S ) = 1
⎪
⎪
⎪
⎪ tr (Q 3 S ) = 1
⎩
S 0
where S ∈ H2N +2 , and the matrices Q −1 , Q 0 , Q 1 , Q 2 and Q 3 are respectively

defined as

W 0 M 0 I − s0
Q −1 = , Q0 = , Q1 = , (22)
00 00 −s †0 s 0 2 − δ
and
I 0 00
Q2 = , Q3 = . (23)
00 01

As a consequence, the optimal solution s (m) of P2 (or equivalently of P2 ) can
be found by employing a semi-definite relaxation (SDR) procedure [9].
4.3 Transmit-Receive System Design: Optimization Procedure
After solving problems Pw and Ps , we can introduce the entire optimization

procedure to iteratively optimize transmit waveform and receive filter for the
output SINR improvement. Precisely, the optimization procedure is summarized
as Algorithm 1. To trigger the recursion, an initial radar code s (0) from which
obtaining the optimal filter w(0) is required; a natural choice is s (0) = s 0 .
10 M. Yuan et al.
Algorithm 1. Polarization Transmit-Receive System Design

Input: σv2 , {σn , n , χn , ρn } , rjk (·, ·), {T (θ)} , s 0 , γ.
Output: A solution (s , w ) to P.
1: set m := 0, s (m) = s 0 , then the receive filter optimizer w (m) is given in Eq. (15),
2
and SINR(m) = g(s (m) , w (m) ) ;
2: do
3: m := m + 1;
4: construct W , M , {Q i }1−1 and {Q i }32 by using (18), (19), (22) and (23) respec-
(m)
tively. Then finding an optimal
radar code s by employing SDR method;
(m)
5: construct the matrix Σc s ;
6: solve Pw (m) finding an optimal receive filter using Eq. (15).
2
7: let SINR(m) = g(s (m) , w (m) ) ;
8: until |SINR(m) − SINR(m−1) | ≤ ζ
9: output s = s (m) and w = w (m) .
5 Numerical Results
In this section, the performance analysis of the proposed algorithm is presented.
We utilize the Georgia Technology Research Institute publicly released dataset
[1] as the foreknowledge of the target. The dataset includes a turntable phase
history data of the T-72 tank measured covering a set of azimuth and elevation
angles by a fully polarimetric X-band radar. More detailed description of the
experimental setup is omitted for the sake of simplicity in this summary and
will be brought in the full version.
5.1 Monotonic Property of the Proposed Method

Figure 1 plots the output SINR versus the number of iterations, for differ-
ent values of the similarity γ ∈ {0.01, 0.10, 0.20, 0.50} and azimuth angles
(1) (2) (3) (4)
θ ∈ {θaz , θaz , θaz , θaz }.
As expected, the plots show the devised method monotonically increases the
lowest output SINR. Moreover, as γ increases, the optimized SINR values are
improved since the feasible set of the optimization problem is enlarged. The
results also highlight the effectiveness of the proposed procedure to provide a
significant performance gain with respect to conventional receiver design, namely
m = 0. In our simulations, SINR gains up to 15 dB, with respect to step zero
of our procedure can be observed for γ = 0.5, even if this is just a potential
value and in real conditions smaller gains could be experienced due to some
inaccuracies in the a priori information.
5.2 SINR Behavior vs. TAA Uncertainty Interval

It is well known that the TIRM is strongly dependent on the TAA in a two-
dimensional scenario, namely the target scattering behavior is sensitive with
(a) (b)
(c) (d)
Fig. 1. The optimized SINR behavior vs. iterations for γ = [0.01, 0.10, 0.20, 0.50]:
(1) (2) (3) (4)
(a) θaz = 27.00◦ ; (b) θaz = 48.70◦ ; (c) θaz = 91.20◦ ; (d) θaz = 133.70◦ .
respect to even a small change of the TAA. Considering (s , w ) as the optimal

transmit code and receive filter pair in case of θ = θ0 , we can define the quantity
SINR(θ) to describe the output SINR response w.r.t. the TAA θ change as
|αT |2 |w † T (θ)s |2
SINR(θ) . (24)
E[|w † C s |2 ] + E[|w † v |2 ]
Obviously, the case θ = θ0 means the real TAA mismatches with the nominal
one, and then a SINR loss may appear in comparison of its value at the position
of θ = θ0 .
Let the TAA uncertainty interval be θ ∈ [25.5◦ , 28.7◦ ] and the nominal TAA
θ0 = 27.0◦ . Figure 2 depicts SINR(θ) for γ = 0, 0.05, 0.1 and 0.5, respectively.
The curves highlight that, the proposed optimization approach (γ = 0.05, 0.1
and 0.5) achieves a significant SINR improvement w.r.t the fixed transmit wave-
form design (γ = 0). Furthermore, even though the optimization design shows
the maximum SINR is obtained at the position θ = θ0 , it exhibits higher and
12 M. Yuan et al.
Fig. 2. SINR(θ) vs. θ in the TAA uncertainty interval θ ∈ [25.5◦ , 28.7◦ ], with the
nominal TAA θ0 = 27.0◦ and γ ∈ {0, 0.05, 0.1, 0.5}.
flatter SINR values over the entire TAA uncertainty set w.r.t the fixed transmit
waveform design, which shows its robustness.
6 Conclusions
Joint optimization of the transmit signal and receive filter for the extended
target and polarimetric radar in the presence of signal-dependent and signal-
independent interference, has been considered. By considering the SINR at the
output of a filter as the relevant figure of merit for optimization, a similarity
constraint has been enforced on the transmit code, in addition to the energy
constraint, aiming at enforcing desirable ambiguity function properties. Owing
to a suitable reformulation of the considered non-convex design problem, an
iterative optimization procedure over the transmit signal and the receive filter
has been devised.
Finally, the proposed method has been validated through experimental
results, showing the capability of the procedure to ensure the significant SINR
gains, with respect to the traditional approach adapting the receiver side only.
We have also analyzed the output SINR behavior against the TAA mismatch,
and the results have shown its robustness.
References
1. https://www.sdms.afrl.af.mil/index.php?collection=gtri
2. Aubry, A., Demaio, A., Farina, A., Wicks, M.: Knowledge-aided (potentially cog-
nitive) transmit signal and receive filter design in signal-dependent clutter. IEEE
Trans. Aerosp. Electron. Syst. 49(49), 93–117 (2013)
3. Chen, C.Y., Vaidyanathan, P.P.: Mimo radar waveform optimization with prior
information of the extended target and clutter. IEEE Trans. Sig. Process. 57(9),
3533–3544 (2009)
4. Garren, D.A., Odom, A.C., Osborn, M.K., Goldstein, J.S., Pillai, S.U., Guerci,
J.R.: Full-polarization matched-illumination for target detection and identification.
IEEE Trans. Aerosp. Electron. Syst. 38(3), 824–837 (2002)
5. Giuli, D.: Polarization diversity in radars. Proc. IEEE 74(2), 245–269 (1986)
6. Karbasi, S.M., Aubry, A., De Maio, A., Bastani, M.H.: Robust transmit code and
receive filter design for extended targets in clutter. IEEE Trans. Sig. Process. 63(8),
1965–1976 (2015)
7. Novak, L.M., Sechtin, M.B., Cardullo, M.J.: Studies of target detection algorithms
that use polarimetric radar data. IEEE Trans. Aerosp. Electron. Syst. 25(2), 150–
165 (1989)
8. Trees, H.L.V.: Detection of Signals-Estimation of Signal Parameters, vol. 4. John
Wiley & Sons Inc., New York (2002)
9. Ai, W., Huang, Y., Zhang, S.: Further results on rank-one matrix decomposition
and its application. Mathematical Programming, January 2008
10. Cheng, X., Aubry, A., Ciuonzo, D., De Maio, A., Wang, X.: Robust waveform and
filter bank design of polarimetric radar. IEEE Trans. Aerosp. Electr. Syst. 53(1),
370–384 (2017)
CRAWL: A Trace Routing Algorithm Based
on Hybrid Two-Layer Topology
Li-ming Zheng1(&), Wen-feng Long1, Yu-Jia Liu1,

and Wei-dong Sun2
1
Department of Message Communication, Armed Police Officer Academy,
Chengdu, China
ygrsun@163.com, 249229503@qq.com, 22561847@qq.com
2
National Key Laboratory for Parallel and Distributed Processing,
School of Computer, National University of Defense Technology,
Changsha, China
11812375@qq.com
Abstract. Data distribution is a key technology for resources convergence and

sharing in distributed environment. To better meet the requirement for real time
data distribution in the dynamic network, a trace routing algorithm named
CRAWL based on the hybrid two-layered topology is put forward. The algo-
rithm contains an overlay topology named CBDLO, upper of which consists of
multiple distributed balanced binary trees corresponding to different properties
and the lower of which is an unstructured topology. CRAWL forwards the data
on the lower unstructured topology in the form of random walk, so that the data
can be sent to the corresponding upper topology entry. It also includes a
matching algorithm named CDM for the parallel matching data properties on the
upper distributed and balanced binary tree and transmitting the matched data to
the nodes that are interested in the data. The experimental results show that the
algorithm can effectively support large-scale data distribution in a dynamical
network, reduce distribution overhead and matching delays.
Keywords: Data distribution Publish/subscribe Message routing

Data matching
1 Introduction
Data distribution has become an important part of distributed application, and its real
time must be improved due to the development of emergency response systems. For
example, in network centric operations [1], military information must be sent from
resource node to the interested target nodes as quickly as possible. The ShakeCast [2]
system of US Geological Survey (USGS) sends the damaged figures of public facilities
to users after the earthquake, and it is to help those victims choose the fastest escape
route.
A publish/subscribe model is a key technology for numerous data distribution
when emergency happens. According to different subscribe models, publish/subscribe
systems can be divided into four kinds based on channels, themes, types, and content

https://doi.org/10.1007/978-3-030-15035-8_2
CRAWL: A Trace Routing Algorithm 15
respectively [3]. The expressive competence of Channel-based publish/subscribe sys-

tems is very weak, and the early ones are usually based on themes [4–6]. Type-based
publish/subscribe systems introduce the type models for target users and filter events in
the light of event types, but the efficiency is relatively low and are seldom put into
practice now. Content-based publish/subscribe systems have a strong expressive
capability with their multiple Boolean Expressions [7–10]. To meet the increasing
requirement of complex subscription, content-based publish/subscribe systems are the
major subject of this research.
On the basis of unstructured topological publish/subscribe systems, nodes could
apply topology randomly to ensure the reliability of links, while the instability of nodes
in the unstructured topology and the random transmission of message could cause
relatively much message redundancy [11, 12]. In the structured-topology-based
publish/subscribe systems, the relation among nodes are usually strictly controlled by
definite algorithm (for example, DHT), by which the location of resources or their
meta-information is sent accurately to corresponding nodes, and structured topology
generally has a relatively low node location delay while the maintenance fee for
topology is relatively high [8, 13, 14].
A trace routing algorithm based on the hybrid two-layer topology is put forward in
this research, on the purpose of supporting fast data distribution. Firstly, a hybrid two-
layer topology CBDLO is presented, which integrates flexibility and expandability of
unstructured topology, accuracy and low cost on information of structured topology, so
as to provide robust, low-cost, highly efficient routing and matching support. Secondly,
CRAWL is put forward to decrease data routing’s information overhead of lower
topology in CBDLO. The experimental results of Peersim simulator show that CBDLO
enables to effectively support data distribution in the large scale dynamic network,
CRAQL algorithm could reduce distribution overhead effectively and CDM algorithm
is able to decrease the matching delay effectively.
2 System Structure
To better take advantage of structured and unstructured topology, and offset their
weaknesses, CBDLO (Content-based Two-Layered Overlay) is presented to address
the problem of composing a content-based fast data distribution topology in the
dynamic network, and it integrates the characteristics of the flexibility and expand-
ability of unstructured topology, accuracy and low overhead on the information of
structured topology.
The lower layer of CBDLO is an unstructured net topology for the maintenance of net
connectivity, whose logic structure is similar to Random Graph. Every node keeps a
certain amount of other nodes as neighbors, which is addressed as View and update
regularly to enable the topology to timely reflect the net connectivity and dynamics.
In CBDLO, the way of random sampling of nodes and updating their views makes
topology preserve the structure of random graph and also maintain the characteristics of
small world model and power law. Meanwhile, every node is equipped with a property
topology APT (Access Point Lookup Table) and cache some information of upper
property topology to support the access of other nodes and the routing when transmit data.
16 L. Zheng et al.
CBDLO classifies nodes with the same subscribe property into clusters. Every
cluster is a property topology. According to the property value distribution, it also
divides the whole property value space into several connective but mutually exclusive
subspace. And every subspace corresponds to one virtual node, then all virtual nodes
are mapped to the real ones so as to make up of a distributed and balanced binary tree
structure.
In CBDLO, when data are generated, they firstly are forwarded through random
walk in the lower topology; after matching with the corresponding property topology,
they enter the upper property topology and begin to transmit data in a definite way.
Two-layered net topology structure is shown in Fig. 1.
upper layer:
property topology structure
Lower layer:
unstructured topology structure
Fig. 1. Two-layered net topology structure
3 Description of CRAWL Algorithm
In the lower-layer topology, data transmit by random walk and look for the matching
property topology on every node through the property topology APT. To avoid circuit
in the process of random walk, CRAWL (Content-based Routing Algorithm with
Labeled Trace) is put forward to record the latest routing by taking along a routing trace
buffer when transmit data. To avert path blocking brought by trace, routing recovery
mechanism is introduced.
3.1 Algorithmic Process

After source nodes generate data, every property of data is decomposed, and a random
walk is started to forward data for each property. So, several properties of a datum
transmit data in a parallel way speed up data transmission. After receiving data, every
node seeks for its corresponding property topology in APT. If it succeeds, data will be
directly forwarded to the corresponding property topology; if not, a neighbor node is
randomly chosen to transmit data (Fig. 2).
Lower random sending routing algorithm

/ / T h e n o d e A r e l e a s e a n e w d a t a . D < M e t a d a t a L i s t ,Tr a c e , C o n t e n t > ;
// The implementaƟon process from A node is shown:
step1. for int i = 0 to k do
begin
d a t a = < M e t a d a t a L i s t [ i ] , Tr a c e , C o n t e n t >
data.trace.add(A);
N’ = select(A.view)
send(data,N’);
end;
step2. //after the node A receives a data, data<metadata,trace,content>
s t e p 2 . 1 f o r i n t j = 0 t o N . A P T. c o u n t d o
begin
if N.APT[j].attribute = data.metadata.attribute then
begin
Send data to node N.APT[j].subscriber;
break;
end;
end;
step2.2 if data.tti > k then
return;
s t e p 2 . 3 i f j = N . A P T. c o u n t t h e n
begin
data.trace.add(N);
N ’ = s e l e c t ( v i e w, N ) a n d N ’ d o e s n ’ t b e l o n g t o N . t r a c e
N = N’;
send(data,N’);
goto step3;
end;
step3. end.
Fig. 2. The random transmission routing of the lower topology
Step 1: data are decomposed into the form of single—property—value; source

nodes are added to Trace, and a neighbor node is randomly selected to forward data.
Step 2 presents the process when a certain node receives attributive data; Step 2.1
shows that the data should at first scan their own APT, if succeed, data are directly
forwarded to the corresponding node and end the routing process; if not, Step 2.2 will
judge whether the data reach the maximum hops. If not, the data will enter Step 2.2 in
which the present node is added to the Trace, and then a neighbor node excluding from
Trace is randomly chosen and then data are sent to this node.
3.2 Path Blocking and Recovery

Path blocking means that in the process of random walk, when data arrive at some node
while its neighbors are all located in the Trace, only to find no way to pass. This
18 L. Zheng et al.
situation is illustrated in Fig. 3. In this figure, node A initiates a random walk along the
path following the arrows, B ! C ! D ! E ! F, but node F is on the edge of net
with a single neighbor node. Path return signifies that after the path has been blocked, a
node from Trace could be randomly chosen and then restart a random walk, but to
avoid being trapped into blocking and path recovery repeatedly, the maximum hop of
each new random walk should be reduced to half.
Fig. 3. Path blocking
4 CDM Algorithm
When data find their upper property topology by means of random transmission in the
lower topology of CBDLO, data begin to accelerate matching distribution. To address
data matching distribution on the upper layer balanced binary tree topology, CDM
(Counter-based Distributed Matching Algorithm) is proposed. The matching data
properties on the upper layer distributed and balanced binary tree are identified through
CDM, then matching data could be sent to nodes interested in such data correspond-
ingly. Then subscribe nodes count the matching properties. If the number of matching
properties surpasses that number, data can be received.
Soon after data arrive at property topology, the only task is to find their matching
nodes, then transmit data. Property topology is a balanced binary tree structure, so its
matching process is actually fulfilled on binary tree, but before and after its trans-
mission on the binary tree, two conversions should be implemented.
The specific matching process is shown in Fig. 4. Step 1 represents that data nodes
should firstly confirm whether matching properties of the received data could be found
in the virtual node list; if they do, looking for virtual nodes including the property value
among virtual nodes is followed; if they do not, the subscriber probably has left, so a
new random walk on the lower topology should be restarted. In order to avoid being
stuck in such situation repeatedly, every start of new random walk should reduce the
maximum hop to half.
Property topology matching algorithm

// the node A receives the matching data
//data=<subject,metadata,trace,content>
step1. bool Found = false;
f o r i = 0 t o A . Vi r t u a l N o d e L i s t . c o u n t d o
begin
i f ( m e t a d a t a . a t t r i b u t e = A . Vi r t u a l N o d e L i s t [ i ] . A t t i b u t e ) t h e n
begin
// Find the matching virtual node
step1.1 Found = true;
s t e p 1 . 2 Vi r t u a l N o d e V N =
F i n d Ta r g e t N o d e ( A . Vi r t u a l N o d e L i s t [ i ] , d a t a ) ;
step1.3 for j = 0 to VN.NodeList.count do
begin
Send(data,VN.NodeList[j];
end
break;
end;
end;
step2. if (not Found) then
being
// Don’t find the matching virtual node
data.MaxHop = data.MaxHop/2;
Restart the new transmission process of Gossip from A node.
The iterative number is MaxHop
end;
step3. end.
Fig. 4. Property topology routing algorithm
5 Experimental Results and Analysis
5.1 Method Realization

Peersim is a P2P simulator, which support P2P system simulation by component-based
development so as to make protocol development easy. Therefore, Peersim is applied to
CBDLO simulation test. For the sake of better examination of the CBDLO capability,
Peersim is utilized to realize topology structure by event driven pattern and the test
results are analyzed accordingly.
5.2 Parameter Settings

We suppose the probability of attribute in subscription obey the Zip-f distribution. For
every subscription, the occurrence probability of its attribute in subscription is l ia .
The general attributes belong to most of subscriptions, and some particular attributes
only belong to special ones. For CBDLO, the number of virtual node is equal to
maximum partition number which is produced by a subscription. Because topological
20 L. Zheng et al.
property is a self-balancing binary tree, the height of the tree is equal to N, and N is the
total number of virtual nodes in the property topology.
5.3 Routing Message Overhead

To display CBDLO capability, another protocol called as Simple realized by Peersim,
to compare with CBDLO. Simple is a plain content-based flooding distribution pro-
tocol. In this simple protocol, topology is similar to the unstructured random graph
topology in the lower topology of CBDLO. Its data transmission firstly begins with the
local property matching. If it matches, it means the present node subscribes the data and
receive that, then data are sent until message achieves the maximum hop. Certainly,
this protocol is simple but its cost is still rather huge, and there’s no guarantee of
covering all nodes interested in data.
At first, presume system’s size is 3000 nodes, view is 10, the max cache of local
subscribe is 100, the max hop is 10, then the test results are as follows. One effective
system requires not only the low cost of message redundancy but also the accuracy of
forwarding data to corresponding nodes. For the effectiveness, effective distribution
message overhead is defined as:
5
x 10
4.5
4 Simple protocol
CRAWL algorithm
3.5
successful matching overhead
2.5
1.5
0.5
0
0 500 1000 1500 2000 2500 3000
total subscription number
Fig. 5. Effective distribution cost changing with the total number of subscribe
As shown in Fig. 5, the 1st point of Simple protocol demonstrates that the total
number of all messages in the system is 68605, and theoretically matched distributed
messages should be 408, while the actual matched messages are only 68. So the
overhead of effective distributive messages is 68605 / (68 / 408) = 411630.
In Fig. 5, CRAWL greatly reduces system’s effective distribution overhead.
Because in the random transmission process, the matching probability increases with
the rising amount of subscription, the effective distribution overhead of Simple protocol
gradually decreases with the increase of subscription. However, the increase of sub-
scription has rather slight influence on CBDLO, because in the simulation process
when the subscription reaches a certain amount, the message generated during a certain
time is still fixed. For CBDLO, CRAWL algorithm is affected only on the first stage.
The overhead on the second stage is fixed which equals the height of property topology
tree while the influence on the first stage is just on the number of property and CBDLO
enables to send data accurately to corresponding nodes. Therefore, the amount of

subscription has little impact on CRAWL algorithm.
Figure 6 demonstrates the relation between effective distributive message overhead
and the maximum hop. Compared with Fig. 5, when hops ascend, the effective dis-
tributive message overhead will decrease more obviously if compared with subscrip-
tion number. And compared with Simple algorithm, the result is very stable. Under the
influence of hops, the effective distributive message overhead decreases rapidly
according to Simple algorithm, but when it arrives at a certain degree, it will never
decrease, otherwise increases sometimes. It is because for Simple algorithm the mes-
sage will rise up according to the increase of hops, but the more the hops, the slighter
influence on its accuracy and the message itself becomes the major overhead.
5
x 10
4.5
Simple
CRAWL
overhead of effective distributedive messages
3.5
2.5
1.5
0.5
10 15 20 25 30 35 40 45 50
max hops
Fig. 6. The relationship between overhead of effective distributed messages and the max hops
4
10
TERA
average overhead of distributed messages
CBDLO
3
10
2
10
1
10 2 3 4 5 6 7 8 9
10 10 10 10 10 10 10 10
nodes number
Fig. 7. Overhead of distributed messages comparison between CBDLO and TERA
Figure 7 illustrates the effective distributive message overhead changes with nodes
by TERA and CBDLO respectively. In the experiments, nodes range from 100 to
1,000,000,000. TERA has 100 themes. According to the experiment of literature, the
random transmission hit rate of TERA lower topology goes up with the increase of
APT. When APT is 50, max hop of random walk is 10, the hit rate could be 90%. In the
large network, the hit rate will not improve obviously at the rise of APT, so APT of
TERA should be 50 taking space and efficiency into consideration.
22 L. Zheng et al.
The hit rate of CBDLO’s lower topology will go up with the rise of APT, but if the
hit rate will be 90% when the max hop is 10, then APT is no less than 100. So APT for
CBDLO is set as 100. Based on the above conditions, the experimental results are
shown in Fig. 7. It shows when network is small, with increasing nodes, the message
overhead difference of TERA and CBDLO is obvious, for which TERA’s broadcast
overhead is not so much in small network and CBDLO has many lower message
overhead; but with the increasing scale of network, TERA’s message overhead
increases quickly while CBDLO increases much slower, which shows the upper
structured topology of CBDLO effectively reduces message overhead. Thus, CBDLO
is better than TERA and more adaptable to the large scale network data development.
5.4 Content Matching Overhead

CDBLO distributes the matching process to different nodes of the network so that the
matching process can work on those nodes at the same time. CRAWL realizes dis-
tribution of data matching filter by two mechanisms. The first one is to decompose the
multiple properties of one data and enables different properties to work on different
topology; even if on the lower topology, several properties work concurrently. The
second mechanism is to divide property’s value space into several connecting but
mutually disjoint subspace.
In experiment, it supposes that subscription obeys Zip-f distribution, and the view
in adjacent node is 20, the value of APT is 20, the maximum iterative step is 30.
Figure 8 illustrates the matching overhead changes with subscription number. Because
Simple protocol is a flooding one, and its vertical scale is multiplied by 0.1. However,
the coefficient of CBDLO’s vertical scale is multiplied by 1/8. The property average is
8, so that is actually the matching overhead generated by the single transmission of the
original node. In Fig. 8, it is still found that when subscription is less than 3000,
CBDLO will decrease at the rise of subscription. When the amount of subscription is
small, APT’s hit rate will be improved according to the rise of subscription, then the
overhead will decrease on the first stage.
220
200
180
160
matching overhead
140
CBDLO/8
Simple/20
120
100
80
60
40
20
500 1000 1500 2000 2500 3000 3500 4000 4500 5000
subscription number
Fig. 8. The relationship between overhead of matching and subscription number

180
160
140
overhead of matching
Simple/8
120 CBDLO/20
100
80
60
40
20
60 70 80 90 100 110 120 130 140 150
number of properties
Fig. 9. The relationship between matching overhead and property number
The changes of matching overhead with the property sum are shown in Fig. 9. The
vertical ordinate is dealt with the same way in Fig. 8. It remains to show the strong
advantages of CBDLO. The change of property number has a very slight influences on
CBDLO and this fully displays the effect of parallel matching.
6 Conclusion
Content-based fast data distribution is a key technology for constructing a large scale
distributive system and of great significance for data-centered application. However,
the construction of efficient and robust fast data distribution system confronts many
challenges, such as heterogeneity and dynamism of wide area network, diversity and
dynamism of data set, and the various customer requirements. Through the research of
fast data transmission in the dynamic network systems, a trace routing algorithm
CRAWL based on hybrid two-layer topology is proposed in this paper, which consists
of CBDLO. Besides, a distributed matching algorithm CDM (Counter-based Dis-
tributed Matching Algorithm) is also developed. The experimental results show that the
algorithm proposed in this paper applies to dynamic network with a strong adaptability,
high efficiency of avoiding circuit, good capability of reducing redundant message
overhead and matching delay, as well as improvement of distribution efficiency.
References
1. Dell-Imagine, R.: JTIDS-an overview of the system design and implementation (Joint
Tactical Information Distribution System). In: Position Location and Navigation Sympo-
sium, San Diego, California, pp. 212–215 (2013)
2. ShakeCast [EB/OL] (2014). http://earthquake.usgs.gov/research/software/-shakecast/
3. Eugster, P.T., Felber, P., Guerraoui, R., et al.: The many faces of publish/subscribe. ACM
Comput. Surv. (CSUR) 35(2), 114–131 (2003)
4. Castro, M., Druschel, P., Kermarrec, A., et al.: SCRIBE: a large-scale and decentralized
application-level multicast infrastructure. IEEE J. Sel. Areas Commun. 20(8), 1489–1499
(2012)
24 L. Zheng et al.
5. Rahimian, F., Girdzijauskas, S., Payberah, A.H., et al.: Vitis: a gossip-based hybrid overlay
for internet-scale publish/subscribe enabling rendezvous routing in unstructured overlay
networks. In: IEEE International Parallel & Distributed Processing Symposium (IPDPS),
pp. 746–757 (2011)
6. Zheng, Z., Wang, Y., Ma, X.: PeerChatter: a peer-to-peer architecture for data distribution
over social networks. Inf.-Int. Interdisc. J. 15(1), 259–266 (2011)
7. Pietzuch, P., Hermes, B.J.: A distributed event-based middleware architecture. In: 22nd
International Conference on Distributed Computing Systems Workshops, pp. 611–618
(2012)
8. Gupta, A., Sahin, O.D., Agrawal, D., et al.: Meghdoot: content-based publish/subscribe over
P2P networks. In: Proceedings of the 5th ACM/IFIP/USENIX international conference on
Middleware (Middleware), pp. 254–273 (2004)
9. Grummt, E.: Fine-grained parallel XML filtering for content-based publish/subscribe
systems. In: Proceedings of the 5th ACM International Conference on Distributed Event-
Based System, pp. 219–228 (2011)
10. Li, M., Ye, F., Kim, M., et al.: A scalable and elastic publish/subscribe service. In: IEEE
International Parallel & Distributed Processing Symposium (IPDPS), pp. 1254–1265 (2011)
11. Cao, F., Singh, J.P.: MEDYM: match-early with dynamic multicast for content-based
publish-subscribe networks. In: Proceedings of the ACM/IFIP/USENIX International
Conference on Middleware (Middleware), pp. 292–313 (2005)
12. Wong, B., Guha, S.: Quasar: a probabilistic publish-subscribe system for social networks. In:
Proceedings of the 7th International Workshop on Peer-to-Peer Systems, pp. 2–7 (2008)
13. Aekaterinidis, I., Triantafillou, P.: PastryStrings: a comprehensive content-based
publish/subscribe DHT network. In: IEEE 26nd International Conference on Distributed
Computing Systems (ICDCS), pp. 23–32 (2006)
14. Baldoni, R., Marchetti, C., Virgillito, A., et al.: Content-based publish-subscribe over
structured overlay networks. In: IEEE 25nd International Conference on Distributed
Computing Systems (ICDCS), pp. 437–446 (2014)
Design and Implementation of IoT
Based Class Attendance Monitoring System
Using Computer Vision and Embedded
Linux Platform
Hasan Salman1, Md Nasir Uddin1, Samuel Acheampong2,

and He Xu2(&)
1
College of Overseas Education, Nanjing University of Posts
and Telecommunications, Nanjing 210023, China
hasansalman@ieee.org, f15040102@njupt.edu.cn
2
School of Computer Science, Nanjing University of Posts
and Telecommunications, Nanjing 210023, China
quajo453@yahoo.com, xuhe@njupt.edu.cn
Abstract. To provide reliable, time-saving and automatic class attendance

system, the concept of Internet of Things (IoT) based class attendance moni-
toring system using embedded Linux platform is presented in this paper. The
study is focused on the design and implementation of face detection and
recognition system using Raspberry Pi. The system takes images of students,
and analyzes, detects and recognizes faces using image processing algorithms,
where the Haar cascade classifier algorithm is implemented to detect faces and
local binary pattern histogram algorithm is used to recognize these faces. After
collecting image processing data, the system generates a final attendance record
and uploads it in a cloud server. The cloud server has been implemented using
python based web framework. The record can be accessed remotely from a user-
friendly, web application using the Internet. Finally, the system is also capable
of sending an email notification with the final record to the teachers and students
in a specific time. Tests and performance analysis were done to verify the
efficiency of this system.
Keywords: IoT Embedded Linux platform Haar Cascade Classifier

Raspberry Pi Class attendance
1 Introduction
The Automated Attendance System is the progress that has taken place in the field of
automation replacing the attendance marking activity traditionally. The traditional
method of attendance marking is very time consuming and becomes difficult when the
strength is high. Automation of attendance system has boundary over the conventional
approach as it saves time and also can be used for security purposes. This system also
helps to prevent fake attendance. So automating the attendance process will increase
the productivity of the class and make good attendance records.

https://doi.org/10.1007/978-3-030-15035-8_3
26 H. Salman et al.
To overcome the disadvantages of the manual system, an automatic attendance

recording system is implemented using Raspberry Pi, which is an embedded Linux
platform. The system is applied on the credit card sized Raspberry Pi board which is
connected to a camera for image capturing with the extended capability of open source
Computer Vision (OpenCV) software which is used for image processing operation
like face detection and recognition. We propose a method that takes the attendance
using recognition of face based on the Internet of Things.
Internet of Things (IoT) enables us to develop a system without human interference
and errors. In other words, IoT is an environment that can transfer data over a network
without the human to human or human to computer interactions. The IoT allows
objects to be sensed and controlled remotely across existing network infrastructure,
creating opportunities for direct integration between the physical world and computer-
based system and it provides more efficiency, accuracy as well as economic benefits.
Keeping the above statement in the forefront, the method includes a Cloud server that is
implemented using Raspberry Pi. The server is used to store the final record of
attendance which can be accessed using a web application. The system also includes an
email notification feature for students and teachers. For developing the system, we have
used Haar Cascade Classifier algorithm to detect faces and Local Binary Pattern His-
togram (LBPH) algorithm to recognize these faces. All needed programs were written
in Python. Thus, the development of IoT based class attendance system provides an
easy way to keep track of student attendance data and monitor their progress and
productivity automatically.
2 Related Work
In this section, we reviewed some related technologies, innovative and previous works
on the topic of attendance monitoring system based application.
The most widely used location-based technologies are GPS, Radio Frequency
Identification (RFID), Bluetooth and so on. Researchers work on these technologies for
improving the services for its accuracy, reliability, and efficiency [1]. A lot of appli-
cation exists in different ways to adequately monitor the attendance of every student.
Usually, attendance management uses student ID for authentication. Many researchers
[2] have designed and implemented an ID-based authentication system. This kind of
system required improvement as IDs can be shared or tampered.
Other attendance monitoring techniques use face detection based student atten-
dance system. Problem with face detection based systems is that individuals have to
carry face detection which is not a cost-effective solution [3]. However, the
improvement of the system has been discovered by M. Wong. By introducing palm-
print which developed an attendance system to record employee attendance [4].
Meanwhile, a similar project has been implemented which applied wireless Iris
recognition attendance management system. Biometric identification security is the
primary concern as if one lost its biometric identification from the database then the
user of the system has to face many challenges. Mohammad et al. have defined location
based time and attendance system, that use location as the key of attendance [5]. The
employees in an organization can be determined with the help of a GPS device if both
Design and Implementation of IoT Based Class Attendance Monitoring System 27
coordinates match then the employee is present in the organization. Which can be
extended for different applications and on all mobile phones.
3 System Architecture
The architecture of our class attendance system based on the IoT is shown in Fig. 1.
Fig. 1. The block diagram of proposed system.
The whole system is implemented on the embedded Linux platform using Rasp-
berry Pi board. Raspberry Pi is used as core processor, complete final processing of
images, data storage, real-time face detection and recognition and data server, after that
the data will be present on the network eventually. The power supply is given to the
Raspberry Pi which is the central part of the proposed system. The Pi camera is
connected to the Raspberry Pi CSI port. The camera captures the images of the students
who are present in the class. Raspberry Pi takes those images as input images and
compares the input images with the existing image dataset. This process happens due to
importing the open CV packages at the initial stage of the development of the system.
After detection and recognition, the system generates an excel sheet containing student
attendance result. All of these processes can be controlled by using a web application.
The system collects Excel file and uploads the data into the cloud server which has
been developed by flask web framework in Raspberry Pi. Real-time data visualization
can be accessed and downloaded directly from the cloud server through the web
application.
As a part of the Internet of Things, it requires network connectivity through WiFi
which is connected with Raspberry Pi. The web application can be used by the internet
and collect attendance result with minimum human intervention. The system is also
capable of sending the result by email to respective teacher and student.
28 H. Salman et al.
4 Proposed Approach
The proffered system management will use three step mechanisms: Dataset creation,
training dataset, testing.
In dataset creation, the system Initialize the camera connected to Raspberry Pi, get
image and user id as input. After converting the image into grayscale, it detects the
face’s images of the students in a specific class, who are present in the classroom and
store it in the dataset.
In training dataset part it initializes LBPH face recognizer and gets faces and ID’s
from dataset folder to train the LBPH face recognizer. After training data, it is stored as
XML or YML file. In testing part the system load Haar classifier, LBPH face recog-
nizer and trained data from XML or YML file. Then capturing the image from a
camera, it converts it into grayscale, detects and predict the face using the above
recognizer. If the system recognizes a student’s face, it will be reported or recorded as
present and all those not in the class or not recognize will be absent.
Then, the system collects these attendance data and generate an Excel file and
upload the file in the cloud server. Finally, it sends email to teachers and students
containing attendance data from the process. The flow chart of the proposed approach
is given in Fig. 2.
Fig. 2. Flow chart of the proposed approach.
The Raspberry Pi is the central part of the project. All the data and algorithm has
been processed and programmed with python language in Raspberry Pi. The Pi camera
has been used to capture photos. A cloud server has been developed by using the flask
web framework of python. To access the class attendance system from smartphone or
PC, a user-friendly web application has been developed using HTML and
Bootstrap. The main aim of this proffered management system: giving a vivid and
accurate attendance record of the students, it prevents human intervention concerning
the attendance record, mailing students and teachers using the easiest way of checking
attendance of students in a specific classroom.
5 Hardware Requirements
5.1 Raspberry Pi 3 Model B
Raspberry Pi is an ARM-based single board computer. The Raspberry Pi 3 Model B is
used in this paper, which is the third generation Raspberry Pi [6]. The operating system
used for Raspberry Pi is Raspbian as it is open source. Raspbian is a Linux-based
computer operating system. It has 40 pins in which 24 are GPIO pins and these pins are
used for general purpose.
5.2 Pi Camera
Pi Camera is a custom designed add-on for Raspberry Pi. This interface uses the CSI
interface designed especially for interfacing to cameras [7]. The sensor has 5
megapixels of resolution and has a fixed focus lens on board.
6 Software and Algorithm
6.1 OpenCV
OpenCV is a computer vision software library. The library has a lot of optimized
algorithms, which can be used in many IoT related sectors including face detection and
recognition. As the libraries of our project, we use the Haar classifier and LBPH face
recognizer.
6.2 Haar Cascade Classifier Algorithm

A Haar Cascade is a classifier used to detect the object for which it has been trained
from the source. The proposed system uses Haar Cascades classifier as a face detection
algorithm [8]. Firstly, to train the Haar cascades classifier, the algorithm needs 300
positive images and negative images. Positive images are images with clear faces
where negative images are those without any faces. Haar cascades are similar to the
convolutional kernel which are shown below in Fig. 3(a).
From the difference of the sums of pixels in the white and black rectangle, each
feature is represented as a single value obtained. For calculating of plenty of features,
all different possible sizes and locations of the classifier are used. The arithmetic
computations become time-consuming as the number of classifiers increases. Instead,
30 H. Salman et al.
the concept of integral image has been used. The integral image is derived by using the
following equation:
IR ðx; yÞ ¼ R iðx0 ; y0 Þ
x0 x
y0 y
Fig. 3. Haar Cascades.
To avoid the complexity of calculation, we use the Adaboost machine learning

algorithm, an OpenCV library to eliminate the classifiers redundancy. A weak classifier
has the probability of 50% of detection. The Sum of weak classifier provides a strong
classifier. Classification takes place in stages, if the selected region fails in the first
stage, it will be discard. The classifiers will not be used on that region which is
discarded. The region which passes all the stages i.e. all strong classifiers is treated as
the detected face. Detected Faces are passed to the face recognition phase.
For face recognition, the Local Binary Patterns histogram algorithm (LBPH) has
been used. This Local binary pattern subjects the detected integral image. Face
recognition is much vulnerable to changes like facial expressions, brightness, and
position. For reducing this problem face preprocessing technique is used. For face
preprocessing, we use histogram equalization. For efficiency we use separate prepro-
cessing which is histogram equalization for left and right face. So histogram equal-
ization is done three times, firstly for the whole face and the other two for side faces [9].
Stages of the cascade classifier [10] are followed according to Fig. 4.
Fig. 4. Stages of the cascade classifier formula

6.3 Python
Python is a high-level programming language which is an open source working for a
general purpose language. It can be embedded into any application and can run on all
current operating systems. Overall Python is suitable as a development language for
customizable applications.
6.4 Numpy
Numpy is a library for scientific computing in Python. It provides a high-performance
multidimensional array object and matrices, and along with an extensive collection of
tools to operate on these arrays. For converting images of our project into multidi-
mensional or 2D-array representation, and conversions from grayscale to color images.
6.5 Flask Framework

Within the Python community, Flask is a very popular web framework [11]. It provides
simplicity and flexibility by implementing a bare-minimum web server, and thus
advertises as a micro-framework. Flask web applications is a simple solution for IoT
based monitoring system. In this paper, it has been used to develop a cloud server in
Raspberry Pi.
6.6 HTML and Bootstrap

HTML is a standard markup language used for creating web pages and web applica-
tions. In the system, it is used for displaying the final result on the web page. On the
other hand, Bootstrap is a popular CSS framework for web development [12]. Boot-
strap is used to make the web application responsive and user-friendly.
7 Experimental Result
The IoT based class attendance monitoring system using an embedded Linux platform
functionally implemented, and results obtained successfully.
Fig. 5. Result of face detection (Data set) and face recognition (Right side).
32 H. Salman et al.
Table 1. System accuracies

Face count Detection rate Recognition rate
Single face 98%+ 96%
Multiple faces 95%+ (88–90)%
Group faces 95%+ (80–88)%
Fig. 6. Class attendance system based on IoT (Left side) and python application (Right side).
Fig. 7. Web application for user interface and result of attendance system.
Figure 5 shows the system create a dataset for face detection and recognition and
the result of multiple face recognition. In Table 1, the system accuracies data are given.
The data has been collected and verified among 30 students. Figure 6 shows com-
pletion of smart attendance system based on IoT and python application. Figure 7
shows the web application for user interface and result of attendance in excel sheet
which can be downloaded from the application using Internet from smartphone or
computer. The user can register and check attendance for smart class attendance by
clicking buttons of the web application. The system uses SMTP protocol for sending
email to the students and teacher in a specific time. All of these features make the IoT
based system efficient enough for taking attendance of any class. This approach is done
so that in future as the number of students increases the system does not fails and can
provide backup and remain accurate and reliable in the long run. Also, to use the most
efficient algorithm for the system with the use of upcoming enhanced algorithms in
future, the accuracy of the system can be enhanced.
8 Conclusion
IoT based class attendance monitoring system is a time-saving, reliable and useful
technique for taking attendance instead of the manual technique. There are many
systems which can be used for managing attendance, but the implementation of face
detection and recognition system with the IoT and Raspberry Pi technology make the
system economy. Since the proposed system gets integrated with the cloud server, it
becomes system independent, fast, provides data storage and accessed quickly. The
web application and email notification make the system user-friendly. In the future, we
will improve the attendance record system according to the following:
(1) There will be an automatic printing of student’s attendance record for the teaching
staff.
(2) To monitor students during an examination.
(3) To trace a particular student in an organization quickly with the help of this
system and can be utilized for several security applications where authentication is
required to access the privileges of the respective system.
Acknowledgement. This work is financially supported by the National Natural Science

Foundation of P. R. China (No.: 61672296, No.: 61602261), Scientific & Technological Support
Project of Jiangsu Province (No.: BE2015702, BE2016185, No.: BE2016777), Postgraduate
Research and Practice Innovation Program of Jiangsu Province (No.: KYCX17_0798).
References
1. Patel, R., Patel, N., Gajjar, M.: Online students attendance monitoring system in classroom
using radio frequency identification technology: a proposed system framework. Int.
J. Emerg. Technol. Adv. Eng. 2(2), 61–66 (2012)
2. Gowri, Ch.S.R., Kiran, V., Rama Krishna, G.: Automated intelligence system for attendance
monitoring with open CV based on internet of things (IoT). Int. J. Sci. Eng. Technol. Res.
(IJSETR) 5(4), 905–913 (2016)
3. Shoewu, O., Olaniyi, O.M., Lawson, A.: Embedded computer-Based lecture attendance
management system. Afr. J. Comput. ICT 4(3), 27–36 (2011)
4. Mani Kumar, B., Praveen Kumar, M., Rangareddy: RFID based Attendance monitoring
system using IOT with TI CC3200 Launchpad. Int. J. Mag. Eng. Technol. Manag. Res. 2(7),
1465–1467 (2015)
34 H. Salman et al.
5. Uddin, M.S., Allayear, S.M., Das, N.C., Talukder, F.A.: A location based time and
attendance system. Int. J. Comput. Theor. Eng 6(1), 1–2 (2014)
6. Grimmett, R.; Raspberry Pi Robotic Projects. 3rd Edn. Packt Publishing (2016)
7. Abaya, W.F., Basa, J., Sy, M., Abad, A.C., Dadios, E.P.: Low cost smart security camera
with night vision capability using Raspberry Pi and OpenCV. In: 2014 International
Conference on Humanoid, Nanotechnology, Information Technology, Communication and
Control, Environment and Management (HNICEM), Palawan, pp. 1–6 (2014)
8. Pasumarti, P., Purna Sekhar, P.: Classroom attendance using face detection and Raspberry-
Pi. Int. Res. J. Eng. Technol. (IRJET) 05(03), 3–5 (2018)
9. Rajkumar, S., Prakash, J.: Automated attendance using Raspberry Pi. Int. J. Pharm. Technol.
(IJPT) 8(3), 16214–16221 (2016)
10. T. M. Inc.: Train a Cascade Object Detector. http://www.mathworks.se/help/vision/ug/train-
a-cascadeobject-detector.html#btugex8
11. Ronacher, A.: Quickstart (2010). http://flask.pocoo.org/docs/0.12/quickstart/#quickstart
12. http://getbootstrap.com/
An Indoor 3D Positioning Technology
Based on NB-IoT
Donghui Xue, He Xu(&), and Peng Li
School of Computer Science,

Nanjing University of Posts and Telecommunications, Nanjing 210023, China
{1018041229,xuhe,lipeng}@njupt.edu.cn
Abstract. The Narrow Band Internet of Things (NB-IoT) for low-power, wide-
coverage, and low-cost requirements solves the problem of massive Internet of
Things device connectivity. NB-IoT supports device connection with large data
throughput and long standby time. It is unmatched by traditional cellular data
network technology and Bluetooth technology. By combining the NB-IoT
positioning with the Lagrangian multiplier improved constrained least squares
localization algorithm and the GSM mobile terminal delay estimation algorithm
in NLOS environment, a feasible indoor three-dimensional positioning method
is proposed. In addition, it reduces the influence of environmental factors on
positioning and improves the accuracy of indoor three-dimensional positioning.
1 Introduction
Based on the concepts of smart city, big data, and cloud computing, the Internet of
Things technology has developed rapidly. Various physical devices are connected to
the Internet to transmit large amounts of data over the Internet to enable monitoring and
management of physical devices. Massive connections of various types of physical
equipment, and the data transmitted are also diverse. The requirements for transmission
have also increased, and various types of transmission equipment have begun to
integrate. Different from traditional cellular data network technology, Bluetooth tech-
nology [1] and other remote transmission technologies. Under such a drive, Narrow
Band Internet of Things (NB-IoT) [2] was released by the 3GPP, which is the wireless
technology standardization organization. As an evolutionary technology of LTE,
NB-IoT not only has a higher transmission rate, but also has a large number of device
connectivity capabilities. Features such as lower latency and lower power consumption
are the top choices for manufacturers, and this accelerates the application of the Internet
of Things in life.
The indoor three-dimensional positioning, as the final need to overcome the
positioning technology, has a very broad application prospects. For example: indoor
navigation, garage positioning, etc. However, when the device is indoors, the receiving
satellite signal is easily interfered by the obstruction and causes a non-line-of-sight
error, which makes it impossible to accurately obtain the positioning data [3]. In
addition, indoor complex electromagnetic interference has become one of the important
factors affecting positioning. Various factors have limited the further development of

https://doi.org/10.1007/978-3-030-15035-8_4
36 D. Xue et al.
current indoor positioning technology. The emergence of NB-IoT solves the problem.
It utilizes the characteristics of NB-IoT, large connection and deep coverage to solve
the problem of indoor receiving satellite signal receiving interference. At the same
time, NB-IoT has low cost and low power consumption. The characteristics make
NB-IoT more extensible and can be widely used in the current Internet of Things
industry.
In order to improve the fusion of technology and improve the indoor positioning
accuracy, this paper uses the Lagrangian multiplier improved constrained least squares
positioning algorithm in the newly proposed NLOS environment to solve the non-line-
of-sight error problem. And the delay estimation algorithm of the GSM mobile station
reduces the dependence of the terminal equipment on the base station and implements
passive positioning with low dependency.
2 The Advantages and Deployment of NB-IoT

2.1 The Advantages of NB-IoT
Among the various types of data collected by the Internet of Things, the obtained
positioning information has its special significance. Many of today’s devices rely on
the location data [4] of the terminal and then further develop the functionality
according to their requirements. Therefore, the positioning information [5] has high
application value. For example, whether a shared bicycle is accessed according to the
planned location, the location of the package during the entire logistics process, the
location of the public infrastructure, and even the location of the livestock, the location
of the livestock, and the like. However, unlike the traditional IoT communication
technology, the bandwidth of the NB-IoT system is 180 kHz, so as long as any current
bandwidth is greater than 180 kHz, the NB-IoT can be flexibly deployed in it. The
design of the narrowband signal of NB-IoT itself can also improve the gain effect when
covering. Considering the low-power carrying capacity, NB-IoT greatly controls the
effective cost of the system during construction and reduces the resources consumed
during development at the cost of data transmission rate and delay. This feature also
enables NB-IoT to meet the needs of IoT devices that require a wide range of coverage
and require mass production.
NB-IoT has the following features [6]:
1. The number of connectable devices is large. NB-IoT allows a connection of about
50, 000 devices in a 200 kHz bandwidth. In the connection state of a large number
of devices, low latency, high processing sensitivity, and optimized network struc-
ture are unparalleled.
2. Low power consumption. Since the NB-IoT terminal only performs measurement
and statistics of positioning data in an idle state, most of the rest of the time is in the
power saving mode in the standby state. Therefore, each NB-IoT module terminal
can theoretically last for ten years, which is sufficient for the current needs of
frequent use without replacement.
An Indoor 3D Positioning Technology Based on NB-IoT 37
3. Coverage is more extensive. The signal coverage of a mobile network depends on

the link budget. The GPRS link budget is 144 dB, the LTE link budget is 142.7 dB,
and the NB-IoT link budget can reach 164 dB. Compared with these, the NB-IoT
link budget can be increased by about 20 dB, which is not LTE and GPRS. A peak
is reached, and the coverage can be increased by 7 times in an open environment. In
an indoor enclosed environment, 20 dB is equivalent to the loss that occurs when
building through the building. Therefore, even under indoor closed conditions,
NB-IoT has more application advantages. The transmission distance of NB-IoT can
also reach 15 km, which satisfies the range of signal coverage under widely used
conditions, and also provides basic conditions for application in more abundant
scenarios.
4. The terminal module costs are low. At the present, the unit price of each NB-IoT
combination terminal module can be controlled within 5 dollars. With the devel-
opment of electronic chip technology and the further promotion and application of
the NB-IoT module, the price of the combined terminal module will continue to
decrease. Low cost is one of the necessary conditions for various enterprises to put
into mass production, and it is also one of the indispensable conditions for NB-IoT
application in various scenarios.
2.2 NB-IoT Deployment Method

The bandwidth of the NB-IoT is 180 kHz, which defines three deployment modes: in-
band deployment, independent deployment, and protection zone deployment [7].
INBAND deployment: In-band deployment occupies one PRB resource in LTE.
Generally, when the in-band deployment mode is adopted, the number of PCIs of
NB-IoT PCI and LTE is set to be the same. Because the number of PCIs in NB-IoT is
504, which is the same as the number in LTE, and the PCI generation mode and
functions are basically the same, the PCI in LTE can be reused to achieve the same
frequency networking, and the traditional GSM is different ways. So on such an
identical basis, the NB-IoT can utilize any resource block in the middle of the LTE
carrier. The in-band deployment mode is shown in Fig. 1(a).
Stand-alone deployment: This deployment method does not need to rely on LTE at
all, and can achieve the effect of decoupling with LTE. The bandwidth of the GSM
channel is 200 kHz, and the NB-IoT is 180 kHz, so the NB-IoT can operate in the
GSM channel, and each 10 kHz is left as a protection on both sides. The independent
deployment mode is shown in Fig. 1(b).
Guard Band deployment: This type of deployment mode does not need to occupy the
band resources in LTE, but utilizes the unused band resources in the LTE edge pro-
tection band, and then occupies 180 kHz resources for the NB-IoT to use. Figure 1(c)
shows how to deploy the protection tape.
38 D. Xue et al.
The NB-IoT deployment modes are compared in terms of bandwidth, spectrum,

terminal power consumption, coverage, and delay. The comparison results are shown
in Table 1:
Table 1. Comparison of deployment methods

Stand-alone Guard-band In-band
Bandwidth Less bandwidth limiting Limited available Less available
factors bandwidth bandwidth
Terminal power Meet terminal power Meet terminal Meet terminal
consumption requirements power requirements power requirements
Coverage The coverage is large Coverage is less The coverage is
than Stand-alone minimal
Delay The delay is the smallest and The delay is greater The delay is the
the transmission efficiency is than Stand-alone largest
the highest
Capacity Downstream capacity is about The downlink The downlink
50,000 capacity is capacity is
approximately approximately
27,000 19,000
2.3 NB-IoT Physical Layer Frame Structure

The NB-IoT physical layer consists of two parts: the frequency domain structure; and
the time domain structure. NB-IoT works in a half-duplex manner, that is, it can only
send data or accept data at the same time. The downlink receiving bandwidth should be
180 kHz according to requirements, and the subcarrier spacing of 15 kHz is adopted.
Therefore, in terms of frame structure and resource unit, it should be kept unchanged
with LTE. The basic unit in NB-IoT is a subframe, and each subframe is 1 ms.
In the frequency domain, NB-IoT occupies a bandwidth of 180 kHz (one PRB),
and the resource block contains 12 subcarriers, and the interval between each subcarrier
is 5 kHz. The structure is shown in Fig. 1(d).
In the time domain: one slot in the NB-IoT has a length of 0.5 ms, each slot
contains 7 OFDM symbols, and two slots constitute one subframe. The structure is
shown in Fig. 1(e).
(a) InBAND deployment (b) Stand-alone deployment
(c) GUARD BAND deployment (d) Structure in the frequency
(e) Structure on the time
Fig. 1. NB-IoT frame
3 Location Algorithm
Target positioning accuracy is an important parameter when positioning with the

NB-IoT narrowband Internet of Things. According to the characteristics of the NB-IoT,
when the signal is transmitted in space during positioning, it arrives at the receiving end
through a plurality of different paths. Different signal components arrive at the
receiving end at different times, resulting in inaccuracies in the estimation of the delay.
And when the signal component is in the path of transmission, when the occlusion is
encountered, the non-line-of-sight error caused by the inability to perform linear
propagation may also have a great influence on the target positioning. Therefore,
optimizing the delay estimation algorithm and eliminating the influence of non-line-of-
sight error will be an important method to improve the positioning accuracy of NB-IoT.
40 D. Xue et al.
Firstly, based on the influence of measurement noise and non-line-of-sight error

correlation, the cost function is established. According to the Lagrangian multiplier
linear constraint condition, the improved constraint model is obtained. In order to
improve the performance of the algorithm, the method of group positioning is used to
solve the final centroid with appropriate nodes to obtain the optimal solution and
reduce the non-line-of-sight property. Secondly, the sampled signal is calculated to
estimate the time of arrival of the signal. Then use the cyclic delay estimation algorithm
to make an accurate estimation of the delay. Finally, according to the barometric
pressure value measured by the module in real time, the height of the target is esti-
mated, and combined with the output position information, the position of the target is
obtained in the three-dimensional environment, and the requirements for indoor three-
dimensional positioning are completed. The location algorithm flow is shown in Fig. 2.
Fig. 2. Location algorithm flow
3.1 Delay Estimation

After comparing with the similar time delay estimation algorithm, this paper decides to
adopt the GSM signal delay estimation algorithm [8]. The algorithm divides the delay
estimation into two steps: First, the signal is sampled, and the sampled signal is
calculated, and the estimation of the arrival time of the signal and the start and end time
of the arrival are initially obtained. Different from the traditional positioning method
using time difference, when the traditional method is used to locate by time difference,
the base station near the monitoring station has a small transmitting power, and the
adjacent receiving base station receives less power after the signal transmission loss,
resulting in a pole. Big error. Second, using the loop algorithm, after the loop esti-
mation operation of time, get more accurate data.
3.1.1 Calculation of Arrival Time

The filtered signal U received by the receiver is sampled to obtain a sampling sequence
with a time interval of T. Where T0 is the initial moment.
3.1.2 Cycle Estimation

After determining the initial sampling time, one of the time slot lengths is taken to
perform a loop estimation to determine the arrival time. Overcome the effect of noise
on the signal through loop operations (Table 2).
Table 2 Comparison of time delay estimation algorithms

GSM delay TOA circumferential Chan algorithm
estimation positioning method
Positioning Three-dimensional Ideal situation Three-dimensional
dimension
Base station No base station Need to rely on base stations Need to rely on base
connection required stations
Feasibility Highest feasibility Lowest feasibility Feasibility changes due
to environmental impact
Positioning High precision It is difficult to get a Low accuracy when non-
accuracy positioning solution in actual line-of-sight error is large
situations.
3.2 Non-Line-of-Sight Error Optimization

The non-line-of-sight error is caused by the increase of the transmission distance due to
the inability of the signal to travel in the middle of the transmission. So in order to get
the position of the target node, we need to eliminate the non-line-of-sight error.
3.2.1 Original CLS Algorithm

To obtain the positional parameters of the measurement target, use the equation for
position and distance:
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
di ¼ ðx xi Þ2 þ ðy yi Þ2 ð1Þ
Where di is the moving distance between the reference node and the target, which
ðxi ; yi Þ is the coordinate of the current position of the target, and i is a constant. Then
square the two sides of the equation to get the current actual distance.
qi ðx2i þ y2i Þ x2 þ y2 2xxi 2yyi ð2Þ
The equal sign is established only in an ideal environment where there is no line-of-
sight error. Therefore, in the real environment, only the greater than the relationship
is taken.
v ¼ 4C 2 AQA A ¼ diagðd12 ; d22 . . .dn2 Þ ð3Þ
Then, based on the values of di obtained in the Eq. (1), the matrix A is constructed,
and the estimated value obtained of v from the result of A is obtained.
42 D. Xue et al.
3.2.2 CLS Algorithm for Lagrange Multiplier [9]

In the formula (1), the matrix A is calculated based on the values of x, y, but in the
original CLS algorithm, x, y is ignored, and the default is x2 þ y2 ¼ r 2 . However, after
the target is affected by non-line-of-sight errors, the equal sign does not hold. There-
fore, according to the influence of non-line-of-sight error, a Lagrangian multiplier is
introduced to establish a new optimal position [10] about v estimation function.
Since the Lagrangian multiplier requires linear constraints, a weighted least squares
estimate is used, and then a Lagrangian function is established based on this:
1
LðX; aÞ ¼ X T HX þ CT X aT ðBXÞ ð4Þ
2
minðY AXÞT v1 ðY AXÞ þ aðX T pX þ qXÞ

ð5Þ
AX Y
According to the Eq. (4), the value of a is brought into the Eq. (5), and the CLS
non-line-of-sight error elimination algorithm based on the Lagrangian multiplier can be
obtained.
3.3 Height Estimation

Here, the height measurement is divided into two steps: first, measuring the air pressure
at the height of the current position of the target; second, processing the air pressure
data, converting it to the current height of the target and outputting.
The barometric pressure acting at the current altitude is equal to the weight of the
vertical air column at the current unit area to the upper boundary of the atmosphere.
The barometric pressure value decreases as the height increases. Therefore, using the
air pressure measurement module in the NB-IoT module group, the air pressure at the
current target position is measured, and the data obtained by the NB-IoT is converted to
obtain the current height of the target. According to the standard pressure height
formula (6):
P1
Z2 ¼ Z1 þ 18400 ð1 þ a tÞ lg ð6Þ
P2
In the above formula, P2 is the pressure at the current position of the target, P1 is
the ground standard reference air pressure, Z1 is the ground height, the initial value is
0, Z2 is the height of the current position of the target, and a and t are constant.
When the target moves upstairs or downstairs, the air pressure at the location is
continuously updated, and then the height is automatically converted according to the
pressure height formula to monitor the current position of the target in real time.
Because time, weather and other factors will also affect the target’s air pressure at the
same position, but the pressure value changes within the acceptable error range, and
after the high pressure formula conversion, the error will be further reduced, so this
article is only for the target The data conversion estimation is performed at the position
when the height conversion is performed.
4 Conclusion
Aiming at the problems of non-line-of-sight error and multipath effect and delay
estimation caused by signal quality caused by measurement noise and obstruction
during transmission in indoor three-dimensional positioning, the Lager in the NLOS
environment is proposed based on NB-IoT. The Lange multiplier improves the least
squares algorithm and the mobile station delay estimation location algorithm. When
solving indoor positioning and receiving signals, it is affected by interference factors
and cannot be accurately positioned. At the same time, it can effectively and accurately
realize the three-dimensional positioning of measuring the target height in indoor
environment.
Acknowledgements. This work is financially supported by the National Natural Science

Foundation of P. R. China (61602261), CERNET Innovation Project (No. NGII20180605),
Scientic & Technological Support Project of Jiangsu Province (Nos. BE2015702, BE2016185,
BE2016777), Postgraduate Research and Practice Innovation Program of Jiangsu Province
(No. KYCX 17_0798, No. KYCX18_0931) and NUPT STITP (No. SZDG2018014).
References
1. Hallberg, J., Nilsson, M., Synnes, K.: Positioning with bluetooth. In: 10th IEEE International
Conference on Telecommunications, pp. 954–958 (2003)
2. Huang, W.: Current status and development trend of NB-IoT low-rate narrowband internet
of things communication technology. Electron. Test (6), 58–29 (2017)
3. Mahmud, M.S., Qaisar, S.U., Benson, C.: Weak GPS signal detection in the presence of
strong signals with varying ralative Doppler and long integration gain. In: IEEE/ION
Position, Location and Navigation Symposium, pp. 1015–1020 (2016)
4. Liu, Y., Yang, Z., Wang, X., Jian, L.: Location, localization, and localizibility. J. Comput.
Sci. Technol. 25(2), 274–297 (2010)
5. Gu, Y., Chen, Y., Liu, J., Jiang, X.: Semisupervised deep extreme learning machine for
Wi-FI based localization. Neurocomputing 166, 282–293 (2015)
6. Wu, X., Shen, S., Zhao, J.: Research on wireless sensor network convergence network based
on NB-IoT technology. Comput. Appl. Softw. 35(6), pp. 1000–386 (2018)
7. Dai, B., Yuan, Y., Yu, Y.: Narrowband Internet of Things (NB-IoT) Standards and Key
Technologies. People’s Posts and Telecommunications Press, Beijing (2016)
8. Zhai, X., Wu, C.: Time delay estimation algorithm for time difference positioning in GSM
mobile stations. J. Comput. Appl. 36, 22–24 (2016)
9. Li, S., Hua, J., Wang, D., Zhou, K., Chen, F., Yu, X.: Improved Constrained least squares
localization algorithm for lagrangian multiplier in NLOS environment. Chin. J. Sens.
Actuators 31(8), 1004–1699 (2018)
10. Zheng, Z., Hua, J., Jiang, B., et al.: A new non-line-of-sight suppression wireless localization
algorithm using the optimization principle. Chin. J. Sens. Actuators 26(5), 722–727 (2013)
Consideration of Implementation
Method for the Cloud Type Virtual Policy
Based Network Management
Scheme for the Specific Domain
Kazuya Odagiri1(&), Shogo Shimizu2, Naohiro Ishii3,

and Makoto Takizawa4
1
Sugiyama Jogakuen University, 17-3 Hosigaokamotomachi Chiksa-ku,
Nagoya, Aichi 464-8662, Japan
kodagiri@sugiyama-u.ac.jp, kazuodagiri@yahoo.co.jp
2
Gakushuin Women’s College, Tokyo, Japan
shogo.shimizu@gakushuin.ac.jp
3
Aichi Institute of Technology Aichi, Toyota, Japan
ishii@aitech.ac.jp
4
Housei University, Tokyo, Japan
makoto.takizawa@computer.org
Abstract. In the current Internet system, there are many problems using
anonymity of the network communication such as personal information leaks
and crimes using the Internet system. This is why TCP/IP protocol used in
Internet system does not have the user identification information on the com-
munication data, and it is difficult to supervise the user performing the above
acts immediately. As a study for solving the above problem, there is the study of
Policy Based Network Management (PBNM). This is the scheme for managing
a whole Local Area Network (LAN) through communication control for every
user. In this PBNM, two types of schemes exist. As one scheme, we have
studied theoretically about the Destination Addressing Control System (DACS)
Scheme with affinity with existing internet. By applying this DACS Scheme to
Internet system management, we will realize the policy-based Internet system
management. In this paper, to realize management of the specific domain with
some network groups with plural organizations, concept of implementation
method applied for this scheme is described.
1 Introduction
In the current Internet system, there are many problems using anonymity of the net-
work communication such as personal information leaks and crimes using the Internet
system.
As a study for solving the problems, Policy Based Network Management (PBNM)
[2] exists. The PBNM is a scheme for managing a whole Local Area Network
(LAN) through communication control every user, and cannot be applied to the Internet
system. In the existing PBNM, there are two types of schemes. The first is the scheme

https://doi.org/10.1007/978-3-030-15035-8_5
Consideration of Implementation Method for the Cloud Type Virtual Policy 45
of managing the whole LAN by locating the communication control mechanisms on

the path between network servers and clients. The second is the scheme of managing
the whole LAN by locating the communication control mechanisms on clients. As the
second scheme, we have studied theoretically about the Destination Addressing Con-
trol System (DACS) Scheme. As the works on the DACS Scheme, we showed the
basic principle of the DACS Scheme, and security function [14]. After that, we
implemented a DACS System to realize a concept of the DACS Scheme. By applying
this DACS Scheme to Internet system, we will realize the policy-based Internet system
management. Then, the Wide Area DACS system (wDACS system) [15] to use it in
one organization was showed as the second phase for the last goal. As the first step of
the second phase, we showed the concept of the cloud type virtual PBNM, which could
be used by plural organizations [16]. In this paper of Sect. 2, motivation and related
research for this study are described. In Sect. 3, the existing DACS Scheme and
wDACS Scheme is described. In Sect. 4 the proposed scheme and implementation of it
are described.
2 Motivation and Related Research
As a study for solving the problem on Internet System, the study area about PBNM
exists. This is a scheme of managing a whole LAN through communication control
every user. Because this PBNM manages a whole LAN by making anonymous com-
munication non-anonymous, it becomes possible to identify the user who steals per-
sonal information and commits a crime swiftly and easily. Therefore, by applying this
policy-based thinking, we study about the policy-based Internet system management.
Fig. 1. Principle in first scheme
In policy-based network management, there are two types of schemes. The first
scheme is the scheme described in Fig. 1. The standardization of this scheme is per-
formed in various organizations. In IETF, a framework of PBNM [2] was established.
Standards about each element constituting this framework are as follows. As a model of
control information stored in the server called Policy Repository, Policy Core Infor-
mation model (PCIM) [3] was established. After it, PCMIe [4] was established by
46 K. Odagiri et al.
extending the PCIM. To describe them in the form of Lightweight Directory Access
Protocol (LDAP), Policy Core LDAP Schema (PCLS) [5] was established. As a pro-
tocol to distribute the control information stored in Policy Repository or decision result
from the PDP to the PEP, Common Open Policy Service (COPS) [6] was established.
Based on the difference in distribution method, COPS usage for RSVP (COPS-RSVP)
[7] and COPS usage for Provisioning (COPS-PR) [8] were established. RSVP is an
abbreviation for Resource Reservation Protocol. The COPS-RSVP is the method as
follows. After the PEP having detected the communication from a user or a client
application, the PDP makes a judgmental decision for it. The decision is sent and
applied to the PEP, and the PEP adds the control to it. The COPS-PR is the method of
distributing the control information or decision result to the PEP before accepting the
communication.
Next, in DMTF, a framework of PBNM called Directory-enabled Network
(DEN) was established. Like the IETF framework, control information is stored in the
server storing control information called Policy Server, which is built by using the
directory service such as LDAP [9], and is distributed to network servers and net-
working equipment such as switch and router. As the result, the whole LAN is man-
aged. The model of control information used in DEN is called Common Information
Model (CIM), the schema of the CIM (CIM Schema Version 2.30.0) [11] was opened.
The CIM was extended to support the DEN [10], and was incorporated in the
framework of DEN.
In addition, Resource and Admission Control Subsystem (RACS) [12] was
established in Telecoms and Internet converged Services and protocols for Advanced
Network (TISPAN) of European Telecommunications Standards Institute (ETSI), and
Resource and Admission Control Functions (RACF) was established in International
Telecommunication Union Telecommunication Standardization Sector (ITU-T) [13].
Fig. 2. Essential principle
However, all the frameworks explained above are based on the principle shown in
Fig. 1. As problems of these frameworks, two points are presented as follows. Essential
principle is described in Fig. 2. To be concrete, in the point called PDP (Policy
Decision Point), judgment such as permission and non-permission for communication
pass is performed based on policy information. The judgment is notified and trans-
mitted to the point called the PEP, which is the mechanism such as VPN mechanism,
router and Fire Wall located on the network path among hosts such as servers and
clients. Based on that judgment, the control is added for the communication that is
going to pass by.
The principle of the second scheme is described in Fig. 3. By locating the com-
munication control mechanisms on the clients, the whole LAN is managed. Because
this scheme controls the network communications on each client, the processing load is
low. However, because the communication control mechanisms need to be located on
each client, the work load becomes heavy.
When it is thought that Internet system is managed by using these two schemes, it is
difficult to apply the first scheme to Internet system management practically. This is why
the communication control mechanism needs to be located on the path between network
servers and clients without exception. On the other hand, the second scheme locates the
communication controls mechanisms on each client. That is, the software for commu-
nication control is installed on each client. So, by devising the installing mechanism
letting users install software to the client easily, it becomes possible to apply the second
scheme to Internet system management. As a first step for the last goal, we showed the
Wide Area DACS system (wDACS) system [15]. This system manages a wide area
network, which one organization manages. Therefore, it is impossible for plural orga-
nizations to use this system. Then, as the next step, we showed the cloud type virtual
PBNM, which could be used by plural organizations in this paper.
Fig. 3. Principle in second scheme
3 Existing DACS SCHEME and wDACS System
3.1 Basic Principle of the DACS Scheme

Figure 4 shows the basic principle of the network services by the DACS Scheme. At
the timing of the (a) or (b) as shown in the following, the DACS rules (rules defined by
the user unit) are distributed from the DACS Server to the DACS Client.
(a) At the time of a user logging in the client.
(b) At the time of a delivery indication from the system administrator.
According to the distributed DACS rules, the DACS Client performs (1) or
(2) operation as shown in the following. Then, communication control of the client is
performed for every login user.
(1) Destination information on IP Packet, which is sent from application program, is
changed.
(2) IP Packet from the client, which is sent from the application program to the
outside of the client, is blocked.
Fig. 4. Basic principle of the DACS scheme
An example of the case (1) is shown in Fig. 4. In Fig. 4, the system administrator
can distribute a communication of the login user to the specified server among servers
A, B or C. Moreover, the case (2) is described. For example, when the system
administrator wants to forbid an user to use MUA (Mail User Agent), it will be
performed by blocking IP Packet with the specific destination information.
In order to realize the DACS Scheme, the operation is done by a DACS Protocol as
shown in Fig. 5. As shown by (1) in Fig. 5, the distribution of the DACS rules is
performed on communication between the DACS Server and the DACS Client, which
is arranged at the application layer. The application of the DACS rules to the DACS
Control is shown by (2) in Fig. 5.
Fig. 5. Layer setting of the DACS scheme

The steady communication control, such as a modification of the destination

information or the communication blocking is performed at the network layer as shown
by (3) in Fig. 5.
3.2 Communication Control on Client

When using communication control on every user and every client, communication
control may conflict. In that case, a priority needs to be given. The judgment is
performed in the DACS Server side as shown in Fig. 6. Although not necessarily
stipulated, the network policy or security policy exists in the organization such as a
university (1). The priority is decided according to the policy (2). In (a), priority is
given for the user’s rule to control communication by the user unit. In (b), priority is
given for the client’s rule to control communication by the client unit. In (c), the user’s
rule is the same as the client’s rule. As the result of comparing the conflict rules, one
rule is determined respectively. Those rules and other rules not overlapping are gath-
ered, and the DACS rules are created (3). The DACS rules are transmitted to the DACS
Client. In the DACS Client side, the DACS rules are applied to the DACS Control. The
difference between the user’s rule and the client’s rule is not distinguished.
Fig. 6. Creating the DACS rules on the DACS Server
3.3 Security Mechanism of the DACS Scheme

In this section, the security function of the DACS Scheme is described. The commu-
nication is tunneled and encrypted by use of SSH. By using the function of port
forwarding of SSH, it is realized to tunnel and encrypt the communication between the
network server and the, which DACS Client is installed in. Normally, to communicate
from a client application to a network server by using the function of port forwarding of
SSH, local host (127.0.0.1) needs to be indicated on that client application as a
communicating server. The transparent use of a client, which is a characteristic of the
DACS Scheme, is failed. The transparent use of a client means that a client can be used
continuously without changing setups when the network system is updated. The
function that doesn’t fail the transparent use of a client is needed. The mechanism of
that function is shown in Fig. 7.
Fig. 7. Extend security function
3.4 The Cloud Type Virtual PBNM for the Common Use Between Plural
Organizations
In this section, after the concept and implementation of the proposed scheme were
described, functional evaluation results are described. In Fig. 8 which is described in
[16], the proposed concept is shown. Because the existing wDACS Scheme realized the
PBNM control with the software called the DACS Server and the DACS client, other
mechanism was not needed. By this point, application to the cloud environment was
easy.
Fig. 8. Cloud type virtual PBNM for the common use between plural organizations
The proposed scheme in this paper realizes the common usage by plural organi-
zations by adding the following elements to realize the common usage by plural
organizations: user identification of the plural organizations, management of the policy
information of the plural organizations, application of the PKI for code communication
in the Internet, Redundant configuration of the DACS Server (policy information
server), load balancing configuration of the DACS Server, installation function of
DACS Client by way of the Internet In the past study [14], the DACS Client was
operated on the windows operation system (Windows OS). It was because there were
many cases that the Windows OS was used for as the OS of the client. However, the
Linux operating system (Linux OS) had enough functions to be used as the client
recently, too. In addition, it was thought that the case used in the clients in the future
came out recently. Therefore, to prove the possibility of the DACS Scheme on the
Linux OS, the basic function of the DACS Client was implemented in this study. The
basic functions of them were implemented by JAVA language.
4 Consideration of Implementation Method

for the Scheme to Manage the Specific Domain
4.1 Concept of the Scheme to Manage the Specific Domain
This proposed scheme is a scheme to manage the plural networks group. In Fig. 9, the
concept is explained. Specifically, as a logical range to manage organization A and
organization B, network group 1 exists. Similarly, as a logical range to manage
organization C and organization D, network group 2 exists. These individual network
groups are existing methods listed in Fig. 9. When plural network groups managed by
this existing scheme exist, those plural network groups are targeted for management by
this proposed method. For example, when user A belonging to org. A in network
group1 uses the network which org. C belonging to network group2 which is a different
network group holds, administrative organization Y for network group2 refers for
policy information of user A for administrative organization X of network group1 and
acquires it. After it, in the form that policy information registered with Network
Group2 beforehand is collated with the policy information, the final policy information
is decided. As a result, the policy information is applied to the client that user A uses in
network group2, and the communication control on the client is performed. When a
user moves plural network groups as well as the specific network group, it is thought
that the PBNM scheme to keep a certain constant management state is realized.
Specific administraƟve organizaƟon X RelaƟon Specific administraƟve organizaƟon Y

Policy InformaƟon Server Policy InformaƟon Server
ApplicaƟon of ApplicaƟon of
the policy informaƟon the policy informaƟon
LAN LAN LAN LAN

or or or or
WAN WAN WAN WAN
Network of the own Network of the own Network of the own Network of the own
organizaƟon organizaƟon organizaƟon organizaƟon
(Org. A) (Org. B) (Org. C) (Org. D)
Network Group 1 Network Group 2
Domain 1
Movement terminal in Org. A
(1)Movement and connecƟon
(2)Usage depending on policy informaƟon
Fig. 9. Concept of the proposed scheme
This scheme consists of the following three factors.

(Factor1) Method of user authentication
(Factor2) Determination method of the policy information
(Factor3) Distribution method of the policy information
As Factor1, the proposed method of user authentication which is suitable of this

method is described in Fig. 10. Because the proposed PBNM method is for the method
to manage the whole Internet system, the proposed user authentication system also has
a distributed system form. For example, when user A belonging to org. A in network
group1 accesses the network of the network group1, the user authentication process is
generated for the user authentication server for the network group1. On the other hand,
when user A belonging to org. A in network group1 accesses the network of the
network group 2, the user authentication process is generated for the user authentica-
tion server for the network group1. The server name with domain name are required as
the information which are necessary for this user authentication. Based on the server
name with domain name which is incorporated in the DACS CL in advance, the first
access for the user authentication server is performed. After that, input of user name
and password is requested. When these pieces of information are sent over the network,
they need to be encrypted and sent by SSL (TLS). That is, the user authentication is
handled by the organization to which the user belongs. The point to be noted here is the
meaning of user authentication. Maybe, in some way, the user authentication server
which is possessed by the network group that the user accesses may need to be used.
However, that is only a complementary measure.
Fig. 10. Concept of the proposed user authentication method
The user authentication referred to here is the one before distribution of DACS
rules as policy information. It is explicitly necessary to distinguish it from user
authentication for permitting network connection as a separate one. Depending on the
implementation, it may be descriptively possible to integrate these two user authenti-
cations as one process, but this is a matter to be considered at the implementation stage.
As Factor2, concept of policy information decision method is described. In Fig. 11,
as an example, the user belonging to the organization A connects the notebook client
computer to another organization in Network Group 2. In this case, the DACS rules as
policy information are extracted from the DACS SV as policy information server which
Specific administraƟve organizaƟon X Specific administraƟve organizaƟon Y
Policy Policy
AuthenƟcaƟon InformaƟon AuthenƟcaƟon
InformaƟon RelaƟon
Server Server Server
Server
LAN LAN LAN LAN

or or or or
WAN WAN (a) WAN WAN
Decision of
Network of the own Network of the own Policy Network of the own Network of the own
InformaƟon
organizaƟon organizaƟon organizaƟon organizaƟon
(Org. A) (Org. B) (Org. C) (Org. D)
Network Group 1 Network Group 2
Domain 1 (b)
ApplicaƟon of
Policy
InformaƟon
User in Org. A User in Org. A
Fig. 11. Concept of Policy Information Decision Method
are located in both network group (Network Group 1 and Network Group 2). After it,
the DACS rules are determined and applied for the DACS CL in the client computer.
When the DACS rules are determined, multiple points are considered as follows.
(Point1) Use of the information service of the organization to which the user
belongs
(Point2) Use of the information service of the organization to which the user
connects
(Point3) Use of the Packet Filtering Service
As the first example, with the user in organization A connected the own client to the
network of network group 2, the user accesses the mail system in own organization. At
that time, to enable the client to access the mail service, the communication must be
permitted by packet filtering setting. In the setting of the mail client in the client, the
server name of the mail system of its own organization is normally specified. As the
result, it becomes possible to access the mail system from the client. In this case, as the
packet filtering setting, the policy information transmitted from the policy information
server of the management organization X is reflected on the client. As the second
example, with the user in organization A connected the own client to the network of
network group 2, the user accesses the mail system of the connected organization, like
the first example, to enable the client to access the mail service, the communication
must be permitted by packet filtering setting. In addition, because the server name of
the mail system of its own organization is normally specified in the setting of the mail
client in the client, it is necessary to change the mail system to be accessed by com-
munication control by NAT. As the policy information for these two communication
controls, the one transmitted from the policy information server of the connected
organization is reflected to the client. After it, it becomes possible to access the mail
system of the connected organization from the client. However, in the case of example
2, the policy information used in example 1 also exists, the policy information used in
the above two examples is set in both management organizations. In this case, the fact
that the policy information is set in the connected organization means that the service
use of the connected organization is permitted.
4.2 Implementation Method for the Scheme to Manage the Specific

Domain
As User Authentication Server, an openldap software is adopted. The openldap is a free
software, and has high affinity for the Internet system. Then, as a database for Policy
Information Server, postgresql or mysql is examined. Both databases has also high
affinity for the Internet system. Here, it is assumed that the postgresql is adopted in this
scheme. As a first point, the method of encrypting communication between User
Authentication Server and client computer is considered. Although various encryption
protocols exist, the communication is encrypted by Transport Layer Security (TLS).
The TLS moves on TCP/IP, and realize secure communication channels between
servers and clients. The TLS is separated from application protocols such as Hyper
Text Transfer Protocol (HTTP) and Simple Mail Transfer Protocol (SMTP), Internet
Message Access Protocol (IMAP), and could be used from those application protocols.
Previously, it is called Secure Sockets Layer (SSL). However, it is called TLS now.
TLS corresponds to a openldap software, and is supported by program language such
as JAVA. That is, TLS is adopted considering affinity with the Internet. As a second
point, JAVA is selected as the implementation language. Because the JAVA is used for
the implementation for many services on the Internet and moves on all operation
systems in the client computers, the JAVA is selected.
From here, detail processes are described. At first, the implementation method for
the DACS Client is explained. As a first process, operating system on the client is
initialized after turning on it. After that, the user logs in the client. The DACS Client is
started immediately thereafter. At the beginning of the starting process of the DACS
Client, the user authentication processes is performed, and input of user name and
password is requested. They are sent to the openldap as the authentication server, and
the password checked is performed. As the result of password check, when password
matched, authentication is permitted. After that process, acquisition processing of
DACS rules as policy information (process1) is performed. In the process1, the DACS
Client connects to the DACS Server which is connected to the organization the user
belongs to. Based on the user name and password which are sent from the DACS
Client, control information for the user is extracted from the postgresql which stored
the policy information. As the result, the DACS rules are created, and they are resent to
the DACS Client. Then, as the result of password check, when password didn’t mat-
ched, authentication is not permitted. Process1 is canceled. After process1 is com-
pleted, process 2 is executed. This is process in the network of the organization the user
is actually connected to. Basically, it is the same as process1 of DACS rules acquisition
process. The connected user authentication server and DACS Server replace the server
existing in the organization’s network to which the user actually connects. However, in
the case of operation in which authentication is performed by the authentication server
at the time of logging in to the OS, user authentication in the processing process2 can
be omitted. As the result of process1 and process2, two kinds of DACS rules gathers on
the DACS Client. When the control information in them overlap, one takes precedence.
In principle, control rules that exist in the DACS Server on the network of the orga-
nization to which the user actually connects are prioritized. However, based on the
judgment of the organization, it becomes also possible to prioritize the control rules of
the organization to which the user belongs. For example, the following two control
rules are prepared. Rule 1 is control rule in the DACS Server in the organization to
which the user normally belongs. Rule 2 is control information in the DACS Server in
the organization to which the user is actually connected.
(Rule 1) Destination IP address and port before conversion: 133.20.25.10:80
Destination IP address and port after conversion: 133.20.25.11:80
(Rule 2) Destination IP address and port before conversion: 133.20.25.10:80
Destination IP address and port after conversion: 133.20.25.12:80
According to the above principle, since the control information of Rule 2 takes
precedence, this information is applied to the DACS Client. As a result, when
launching the Web browser and accessing the 133.20.25.10: 80 Web page, actually, the
Web page of 133.20.25.12: 80 is accessed. However, when setting information indi-
cating that rule 1 is prioritized, rule 1 takes precedence and is applied to the DACS
Client.
5 Conclusion
In this paper, implementation Method for the Cloud Type Virtual Policy Based Net-
work Management Scheme for the Specific Domain is described. Considering affinity
with the Internet system, the distributed authentication method was examined. In the
future, implementation will be carried out with details of the authentication method
stuffed up. In the future, implementation will be carried out based on this paper.
Acknowledgments. This work was supported by the research grant of JSPS KAKENHI Grant
Number 17K00113, and the research grant by Support Center for Advanced Telecommunications
Technology Research, Foundation (SCAT). We express our gratitude.
References
1. Cerf, V., Kahn, E.: A protocol for packet network interconnection. IEEE Trans. Commn.
COM-22, 637–648 (1974)
2. Yavatkar, R., Pendarakis, D., Guerin, R.: A Framework for Policy-based Admission Control.
IETF RFC 2753 (2000)
3. Moore, B., et al.: Policy Core Information Model – Version 1 Specification. IETF RFC 3060
(2001)
4. Moore, B.: Policy Core Information Model (PCIM) Extensions. IETF 3460 (2003)
5. Strassner, J., Moore, B., Moats, R., Ellesson, E.: Policy Core Lightweight Directory Access
Protocol (LDAP) Schema. IETF RFC 3703 (2004)
6. Durham, D., et al.: The COPS (Common Open Policy Service) Protocol. IETF RFC 2748
(2000)
7. Herzog, S., et al.: COPS usage for RSVP. IETF RFC 2749 (2000)
8. Chan, K., et al.: COPS Usage for Policy Provisioning (COPS-PR). IETF RFC 3084 (2001)
9. CIM Core Model V2.5 LDAP Mapping Specification (2002)
10. Wahl, M., Howes, T., Kille, S.: Lightweight directory access protocol (v3). IETF RFC 2251
(1997)
11. CIM Schema: Version 2.30.0, 2011
12. ETSI ES 282 003: Telecoms and Internet converged Services and protocols for Advanced
Network (TISPAN); Resource and Admission Control Subsystem (RACS); Functional
Architecture, June 2006
13. ETSI ETSI ES 283 026: Telecommunications and Internet Converged Services and
Protocols for Advanced Networking (TISPAN); Resource and Admission Control; Protocol
for QoS reservation information exchange between the Service Policy Decision Function
(SPDF) and the Access-Resource and Admission Control Function (A-RACF) in the
Resource and Protocol specification, April 2006
14. Odagiri, K., Yaegashi, R., Tadauchi, M., Ishii, N.: Secure DACS scheme. J. Network
Comput. Appl. 31(4), 851–861 (2008)
15. Odagiri, K., Shimizu, S., Takizawa, M., Ishii, N.: Theoretical suggestion of policy-based
wide area network management system (wDACS system part-I). Int. J. Networked Distrib.
Comput. (IJNDC) 1(4), 260–269 (2013)
16. Odagiri, K., Shimizu, S., Ishii, N., Takizawa, M.: Suggestion of the cloud type virtual policy
based network management scheme for the common use between plural organizations. In:
Proceedings of International Conference on International Conference on Network-Based
Information Systems (NBiS-2015), pp. 180–186, September 2015
A Model of Virtual Machines to Support
Storage Processes
Naohiro Ogura1(B) , Dilawaer Duolikun1 , Tomoya Enokido2 ,

1
Hosei University, Tokyo, Japan
naohiro.ogura.2b@stu.hosei.ac.jp, dilewerdolkun@gmail.com,
2
Rissho University, Tokyo, Japan
eno@ris.ac.jp
Abstract. Virtual machines are widely used to support applications

with virtual computation service of computer resources in a cluster of
servers. Application processes on a virtual machine can take usage of
not only CPUs but also storage resources like files and databases inde-
pendently of heterogeneity and location of each host server. In addition,
virtual machines can migrate from a host server to a guest server in
the live manner. In this paper, we consider storage processes which read
and write data in files of storages like HDD. We propose a model for
performing storage processes on virtual machines and migrating virtual
machines supporting files from host servers to guest servers. Here, every
file used by application processes is stored in a storage server and every
virtual machine supporting files resides on the storage server. If a stor-
age process using a file is issued, a virtual machine supporting the file is
found in servers. Virtual machines migrate to working servers, and stor-
age processes on virtual machines are performed. If a virtual machine
supporting read-only files is overloaded, a clone of the virtual machine
is created on the storage server and new processes reading the file is
performed on the clone.
Keywords: Virtual machines · Live migration · Storage processes ·

Storage servers · Working servers · Migration · Clones
1 Introduction
Applications on clients use computation resources supported by a cluster of
servers like cloud computing systems [14]. A cluster of servers provides applica-
tions with virtual computation services through virtual machines [3]. That is,
applications can take usage of computation resources like CPUs and storages
like HDD through virtual machines without being conscious of heterogeneity
and location of each server in a cluster. In addition, virtual machines migrate
from a host server to a guest server in the live manner [3]. That is, application
https://doi.org/10.1007/978-3-030-15035-8_6
58 N. Ogura et al.
processes performed on a virtual machine can migrate to a guest server without

suspending the application processes. Algorithms to migrate virtual machines
from a host server to a guest server, where application processes only use CPUs,
are so far proposed [8], where the electric energy consumed by the servers can
be reduced. In papers [5–7], a group of multiple virtual machines migrate from
a host server to a guest server. In addition, virtual machines are dynamically
resumed and suspended as the number of processes performed increases and
decreases, respectively [8,9]. Here, the total electric energy consumption of all
the servers and the average execution time of application processes are shown
to be reduced compared with non-migration algorithms in the evaluation.
Power consumption models are proposed to give the electric power consumed
by a whole server to perform application processes in the macro-level approach
[9]. In the MLPCM (multi-level power consumption) model [12], the power con-
sumption of a server depends on the number of active CPUs, cores, and threads.
In addition, the computation models of a server to perform application processes
are proposed [9,12], which gives the estimated execution time of each process to
be performed on the server.
In our previous studies [4–8], computation processes, i.e. application pro-
cesses which use only CPUs are considered. Application processes like Web and
database applications use not only CPUs but also files on storage driver. In this
paper, we consider storage processes which read and write data in files of storage
drives like HDD. Even if a virtual machine migrates from a host server to a guest
server, files read and written by processes do not migrate to the guest server.
The files have to be transferred to the guest server by the protocol SFTP [1].
It takes time to transfer every file used by application processes on the virtual
machine. Through experiments, we make clear the execution time of each storage
process on a virtual machine in this paper. We propose a model for performing
storage processes on virtual machines and migrating virtual machines. Here, a
cluster includes two types of servers, storage servers and working servers. Files
manipulated by storage processes are stored in storage servers. Files created on
application processes on a virtual machine can be used only through the vir-
tual machine. Every virtual machine is initially created on a storage server. If a
client issues a cluster an application process which uses files, one virtual machine
holding the files in the storage server migrates to a working server and the appli-
cation process is performed on the virtual machine. An application process can
migrate to another working server while remotely reading and writing files in
the storage server but files do not migrate from the storage server. For an appli-
cation process which uses read-only files, if a virtual machine holding the files
is more loaded, a clone [2] of the virtual machine is created and the process is
performed on the clone. If a host server of a virtual machine is congested, the
virtual machine migrates to another working server.
In Sect. 2, we present the system model. In Sect. 3, we measure the execution
time of storage processes on virtual machines and time to create clones. In Sect. 4,
we present a model for performing processes on virtual machines and migrating
virtual machines and making clones of virtual machines.
A Model of Virtual Machines to Support Storage Processes 59
2 System Model
2.1 Virtual Machines
A cluster is composed of servers s0 , s1 , ..., sm (m ≥ 0) and provides applica-

tions with virtual computation service on computation resources like CPUs and
storages through virtual machines. Thus, a cluster supports a set V M of virtual
machines vm1 , ..., vmv (v ≥ 1) each of which resides on a server of the servers
s0 , s1 , ..., sm . Each virtual machine vmk resides on a server st which is referred
to as host server of the virtual machine vmk . Application processes on a vir-
tual machine can use computation resources without being conscious of which
server supports the computation resources, i.e. independently of heterogeneity
and location of each server. Furthermore, virtual machines can migrate from a
host server to a guest server in the live manner [3]. That is, application processes
on a virtual machine can migrate from a host server to a guest server without
suspending the application processes. A cluster of a virtual machine vmk can be
created on the host server of vmk . In this paper, a process means a storage type
of application process which manipulates files.
A virtual machine vmk is composed of virtual memory mk and virtual storage
vsk . Processes are realized in the virtual memory mk . Suppose a virtual machine
vmk migrates from a host server sh to a guest server sg . Here, the virtual memory
mk of the virtual machine vmk on the host server sh is transferred to the guest
server sg . Only the directory on files in the virtual storage vsk is transferred
to the guest server sg but files themselves pointed by the directory are not
transferred from the host server sh to the guest server sg . Files have to be
transferred from the server sh to the guest server sg by SFTP [1], in order to
migrate the files of the virtual machine vmk . A file which created by a process
on a virtual machine is physically stored on the host server.
First, a virtual machine vmk is created on a host server sh as shown in
Fig. 1. The server sh where the virtual machine vmk is created is referred to as
the home server of the virtual machine vmk . A virtual machine vmk migrates
from a host server sh to another guest server sg . An application process pi also
can create a file fi on the virtual machine vmk hosted by a server sh . Here, the
file fi is physically realized on the host server sh wherever the virtual machine
vmk resides in the cluster S. A server where a file fi exists is a home server
of the file fi . Application processes on the virtual machine vmk read and write
data in the file fi which is stored in the home server sh of fi . Each process pi on
the virtual machine vmk on a server sg remotely reads and writes data in the
file fi of the home server sh . The file fi can be manipulated by processes on the
virtual machine vmk . Thus, a virtual machine vmk where a file fi is created is
a home virtual machine of the file fi .
We consider read-only and updatable files. Virtual machines which hold only
read-only files are static ones. Virtual machines which hold some updatable files
are dynamic ones.
60 N. Ogura et al.
Fig. 1. Migration of a virtual machine with a file.
2.2 Clones
A clone vml of a virtual machine vmk can be created on the host server sh of
the virtual machine vmk . Here, the virtual machine vmk is a parent of the clone
vml . There are two types of clones, full and linked clones [2]. If a full clone vml
is created from the parent virtual machine vmk , not only the virtual memory
mk but also files in the virtual storage vsk of the virtual machine vmk are copied
to the full clone vml . Full clones of a virtual machine are independent of one
another.
A linked clone vml shares virtual storage with a snapshot of a parent virtual
machine vmk . When the clone vml is created, the virtual storage of the clone
vml is a snapshot of vml . Files updated by the parent vmk do not affect the
clone vml and files updated by the clone vml do not affect the parent vmk ,
either. It takes shorter time to create a linked clone than a full clone.
In this paper, we consider a full clone.
3 Experiments
We consider a pair of homogeneous servers sh and sg as shown in Table 1. The

servers sh and sg are equipped with the same CPU, memory (8 [GB]), and
HDD (1 [TB]). The servers sh and st are interconnected in a 1 [Gbps] local area
Table 1. Servers.
Servers sh sg
CPU Intel Core i5-8400 Intel Core i5-8400
Memory [GB] 8 8
HDD [TB] 1.0 1.0
OS CentOS 7.4 CentOS 7.4
Fig. 2. Servers.
network (LAN) as shown in Fig. 2. The same operating system Linux CentOS
7.4 [1] is installed in a pair of the servers sh and sg .
A virtual machine vmk is realized in KVM [3] and supports CentOS 7.4 [1].
The server sh is a home server of the virtual machine vmk . The virtual machine
vmk provides the virtual memory mk and the virtual storage vsk . A virtual
machine vmk migrates from a host server sh to a guest server sg by issuing a
command “virsh migrate” on the host server sh . In addition, a clone vml of a
virtual machine vmk can be created on the host server sh of vmk .
We consider a storage process pi which reads data in a file fi as shown in
Fig. 3. The file fi is created by an application process on a virtual machine vmk
in the server sh . The virtual machine vmk is a provider of the file fi . In the
process pi , the file fi is first opened by an open system call. Here, the size
of the file fi is 5, 15, 20 [GB]. Then, data in the file fi is iteratively read by
read system calls. Data of 1,024[B] is read by one read system call. Then the
file fi is closed by a close system call. Here, every file fi read by a process pi
is different from the other processes and is the same in size. At time stime, a
process pi starts. Before performing the process pi , a pair of the memory buffer
of the virtual machine vmk and the host server are flushed. Time stime when the
open system call is issued and time etimei when the close system call is issued
by a process pi are obtained in nanosecond order by the gettimeofday system
call [13]. The execution time of the process pi is defined to be etimei − stimei
[sec]. V Tk,t stands for the average execution time of the processes pi on a virtual
machine vmk of a server st . HTh means the average execution time of the process
pi which is directly performed on the server sh .
1. First, we measure a pair of the average execution time HTh and V Tk,h of a
process pi on the server sh and on the virtual machine vmk of the server sh ,
respectively [Fig. 3(1)].
2. Next, the virtual machine vmk on the home server sh migrates to the guest
server sg . Then, after the virtual machine vmk migrates from the server sh
to the server sg , the process pi starts to manipulate the file f . The execution
time V Tk,g is measured [Fig. 3(2)].
62 N. Ogura et al.
Fig. 3. Execution time of a process.
3. The virtual machine vmk migrates from the server sh to the server sg and
then migrates back to the server sh in the live manner. Here, the process pi
is performed in the migrations and end on the server sh . Here, the execution
time V Tk,hgh is measured [Fig. 3(3)].
Figure 4 shows the average execution time HTh , V Tk,h , V Tk,g , and V Tk,hgh
[sec] of the process pi for size of the file fi . HTh is one sixth to one seventh of
the others. It takes six to seven times longer to perform a storage process on a
Fig. 4. Average execution time.

virtual machine than a server. On the other hands, the difference between V Tk,h ,
V Tk,g , and V Tk,hgh are small, several percentage, independently of which server
holds the virtual machine vmk .
Fig. 5. Time to create a full clone with a file f .
Next, we measure time to create a full clone of a virtual machine. Figure 5

shows how long it takes time to create a full clone of the virtual machine vmk on
a host server sh for size of the file f . Time to create a full clone linearly increases
as the size of the file f increases. If a virtual machine vmk holds no file, i.e. file
size 0 in Fig. 5 it takes about 30 [sec] to create a full clone vml of the virtual
machine vmk . If the size of the file f are 10 [GB] and 20 [GB], the creation time
is about 200 [sec] and 400 [sec], respectively.
4 A Model of Storage Applications on Virtual Machines
We propose a model of a virtual machine where storage processes are performed.

A cluster S is composed of a storage server s0 and working servers s1 , ..., sm
(m ≥ 1) which are interconnected in networks as shown in Fig. 6. A set of virtual
machines vm1 , ..., vmv (v ≥ 1) are supported to applications in the cluster S.
The storage server s0 supports application processes on working servers with
storage resources.
Initially, every virtual machine vmk is created and resides on a storage server
s0 . The storage server s0 is the home server of every virtual machine vmk (k =
1, ..., m). Memory image mk of every virtual machine vmk is copied to every
working server st (t = 1, ..., m) so that the virtual machine vmk can migrate to
any working server st anytime. Files which an application process reads, writes,
64 N. Ogura et al.
Fig. 6. Storage server and working servers.
and creates exist on the storage server s0 . The file fi can be manipulated only
through the virtual machine vmk .
A virtual machine vmk is dormant if no process is performed. Every dormant
virtual machine vmk is on the storage server s0 . A virtual machine where at least
one process is performed is active.
Suppose a client issues a process pi to a cluster S. The process pi is performed
as follows.
1. If the process pi uses an existing file fi , a home virtual machine vmk of the
file fi is found in the cluster S. Then, the process pi is issued to the virtual
machine vmk , which may be on a working server st .
2. If the process pi uses a file which does not exist in the cluster, one dormant
virtual machine vmk is taken on the storage server s0 . First, one working
server st is selected. Then, the virtual machine vmk migrates to the selected
server st . The process pi is performed on the virtual machine vmk of the host
server st .
3. An active virtual machine vmk on a working server st migrates to another
working server su if more number of processes are performed on the server st
than some value.
4. Suppose a virtual machine vmk holds a read-only file fi , i.e. vmk is static. If
more number of processes are performed on the virtual machine vmk , a clone
of the virtual machine vmk is created on the storage server s0 .
5. If a virtual machine vmk gets idle, i.e. every process terminates, the virtual
machine vmk migrates to the storage server s0 and gets dormant.
At step 1, a working server st whose estimated electric energy to be consumed

by application processes on the virtual machine vmk is minimum.
Suppose a process pi which reads a read-only file fi is issued to a cluster S.

A home virtual machine vmk holding the file fi is found in the storage server.
Initially, a clone vmk of the virtual machine vmk is created on the storage server
s0 . The process pi is performed on the clone vmk and migrates to some working
server st . If there is a clone vmk of vmk on some working server st , the process pi
is issued to the clone vmk . Here, if the clone vmk is congested, a new clone vmk
of the virtual machine vmk is created on the storage server s0 and the process pi
is issued to the clone vmk . Then, the clone vmk migrates to a working server su .
5 Concluding Remarks
In this paper, we discussed storage processes which create, read, and write files
and are performed on virtual machines. Storage processes read and write data in
files which are stored in the storage server. We proposed the model for performing
storage processes on virtual machines and migrating virtual machines. Here,
every virtual machine is initially on a storage server where every file used by the
virtual machine is stored. If a process which uses a file fi is issued to a cluster, a
virtual machine vmk holding the file fi is found. If an idle virtual machine vmk
is on a storage server, one working server is selected and one virtual machine
vmk migrates to the working server. The process is performed on the virtual
machine. If a virtual machine vmk supports only read-only files, a clone of the
virtual machine vmk is created to perform new storage processes. An active
virtual machine migrates to another working server. If every process terminates
on a virtual machine, the virtual machine backs to the storage server.
We are designing an algorithm to select a working server to which a virtual
machine migrates so that the electric energy consumption of servers [9–12] can
be reduced in a cluster.
Acknowledgements. This work was supported by JSPS KAKENHI grant number

15H0295.
References
1. The centos linux distribution (centos linux). https://www.centos.org/
2. Cloning virtual machines for kernel based virtual machine. https://access.
redhat.com/documentation/ja-jp/red hat enterprise linux/7/html/virtualization
deployment and administration guide/cloning virtual machines
3. A virtualization infrastructure for the linux kernel (kernel-based virtual machine).
https://en.wikipedia.org/wiki/Kernel-based Virtual Machine
4. Duolikun, D., Nakamura, S., Enokido, T., Takizawa, M.: An energy-efficient
dynamic live migration of multiple virtual machines. In: Proceedings of the 21st
International Conference on Network-Based Information Systems (NBiS-2018), pp.
87–98 (2018)
5. Duolikun, D., Watanabe, R., Enokido, T., Takizawa, M.: Asynchronous migration
of process replica in a cluster. In: Proceedings of the 11th International Confer-
ence on Broadband and Wireless Computing, Communication and Applications
(BWCCA-2017), pp. 161–172 (2017)
66 N. Ogura et al.
6. Duolikun, D., Watanabe, R., Enokido, T., Takizawa, M.: An Eco algorithm for
dynamic migration of virtual machines in a server cluster. In: Proceedings of
the 20th International Conference on Network-Based Information Systems (NBiS-
2017), pp. 42–54 (2017)
7. Duolikun, D., Watanabe, R., Enokido, T., Takizawa, M.: Energy-aware dynamic
migration of virtual machines in a server cluster. In: Proceedings of the 11th
International Conference on Complex, Intelligent and Software Intensive Systems
(CISIS-2017), pp. 70–81 (2017)
8. Duolikun, D., Watanabe, R., Enokido, T., Takizawa, M.: An Eco migration algo-
rithm of virtual machines in a server cluster. In: Proceedings of IEEE the 32nd
(AINA-2018), pp. 189–196 (2018)
9. Enokido, T., Aikebaier, A., Takizawa, M.: A model for reducing power consumption
in peer-to-peer systems. IEEE Syst. J. 4(2), 221–229 (2010)
10. Enokido, T., Aikebaier, A., Takizawa, M.: An extended simple power consump-
tion model for selecting a server to perform computation type processes in digital
ecosystems. IEEE Trans. Ind. Inf. 10(2), 1627–1636 (2014)
11. Enokido, T., Takizawa, M.: Power consumption and computation models of virtual
machines to perform computation type application processes. In: Proceedings of
the 9th International Conference on Complex, Intelligent, and Software Intensive
Systems (CISIS-2015), pp. 126–133 (2015)
12. Kataoka, H., Duolikun, D., Enokido, T., Takizawa, M.: Multi-level computation
and power consumption models. In: Proceedings of the 18th International Confer-
ence on Network-Based Information Systems (NBiS -2015), pp. 40–47 (2015)
13. Misawa, A., Tatani, F., Kawaguchi, T.: Linux Command Book for Beginners. SB
Creative (2015)
14. Rafaels, R.J.: Cloud Computing: From Beginning to End. Create Space Indepen-
dent Publishing Platform (2015)
Energy Efficient Scheduling
of Smart Home
Sajjad Khan1 , Zahoor Ali Khan2 , Nadeem Javaid1(B) ,

Sahibzada Muhammad Shuja1 , Muhammad Abdullah1 , and Annas Chand3
1
COMSATS University Islamabad, Islamabad 44000, Pakistan
nadeemjavaidqau@gmail.com
2
Computer Information Science, Higher Colleges of Technology, Fujairah 4114, UAE
3
COMSATS University Islamabad, Abbotabad 22010, Pakistan
http://www.njavaid.com
Abstract. Recently a massive increase in the demand of energy has

been reported in residential, industrial and commercial sectors. Tra-
ditional Grid (TG) with the aging infrastructure is unable to address
the increasing demand problem. Smart Grid (SG) enhanced the TG by
adopting information and communication based technological solutions
to address the increasing electricity demand. Smart Home Energy Man-
agement System (SHEMS) plays an important role in the efficacy of SG.
In this paper, an Improved Algorithm for Peak to average ratio Reduc-
tion (IAPR) in SHEMS is developed. To validate the effectiveness of the
IAPR, comparison is made with the renowned meta-heuristic optimiza-
tion approaches namely Strawberry Algorithm (SA) and Salp Swarm
Algorithm (SSA) using two different pricing scheme. It is illustrated by
simulations results that the IAPR reduced the PAR to a greater degree
as compare to SA and SSA.
Keywords: Smart Grid · Optimization techniques ·

Salp Swarm Algorithm · Strawberry Algorithm
1 Introduction
Over the last few decades, a substantial growth has been reported in the demand
of electricity. To cope with the increasing electricity demand problem, the power
generation companies focused on the installation of new energy generators. How-
ever, installing new generation sources is neither an economical nor environment
friendly option. It is due to the fact that these sources mostly operate by fos-
sil fuels like natural gas, oil and coal. It is believed that these sources emit
carbon dioxide into the atmosphere which causes global warming. An alter-
nate solution to the intensifying electricity demand is the efficient utilization
of existing resources. To efficaciously utilize the existing resources Traditional
Grid (TG) is experiencing numerous challenges such as centralized generation
system, high power losses, aged infrastructure and wired technology etc. One of
https://doi.org/10.1007/978-3-030-15035-8_7
68 S. Khan et al.
the most important challenge is the reliability and sustainability of power grid.
The increasing energy demand in industrial, residential and commercial sectors
has a high influence on the grid reliability, as a result power failures or blackout
occurs frequently [1]. The major factors that affects the reliability of TG are the
lack of communication and inefficacious load management capability. Therefore,
to make the power system more reliable and sustainable, the infrastructure of
Smart Grid (SG) is introduced.
SG is an electric grid which is revolutionized by amalgamating Information
and Communication Technology (ICT) based solutions [2]. These ICT based
solutions enable SG to have a bidirectional communication between the con-
sumers and the utility. This is achieved by amalgamating the Advanced Meter-
ing Infrastructure (AMI) and smart appliances in the existing power system.
AMI enables the consumers to learn their real time energy usage and its respec-
tive cost. The advantages of SG over TG are: reliable grid operations, dis-
tributed energy production, Demand Response (DR) strategies, Smart Meter
(SM), Demand Side Management (DSM) and Smart Scheduler (SS) [3]. DSM
plays a key role in enhancing the efficiency of SG to improve the reliability of
power system. DSM refers to the strategies adopted by the utility companies
to motivate consumers to actively participate in efficacious energy utilization
[4]. For this purpose, incentive is provided to the consumers so that energy is
efficaciously utilized. In return, the consumers have to minimize or shift the
operational time of the electric appliances to other time slots (Off-peak hours).
DR aims to persuade the consumers to modify their energy utilization pattern
in distinct hours (On-peak hours). For this purpose two types of DR policies
are introduced. These policies are incentive based DR policy and price based
DR policy. In the incentive based DR policy, consumers agree to minimize their
energy utilization in order to get some incentives from the utility. Incentive to
the consumers is given in the form of minimized electricity bill. In price based
DR scheme numerous pricing schemes are used across the world to encourage
the consumers to alter their energy utilization pattern. The well known pricing
schemes are Critical Peak Pricing (CPP), Inclined Block Rate Pricing (IBRP),
day ahead Time of Use Pricing (ToUP) and Real Time Pricing (RTP) [5].
Some of the common objectives of SG are cost minimization, optimal load
management, the reliability and sustainability of power system and Peak to
Average Ratio (PAR) reduction. In order to achieve the conflicting objectives
researchers have implemented many meta-heuristic and Linear Programming
(LP) techniques [6]. In this paper, we present an Improved Algorithm for PAR
Reduction (IAPR) in the environment of a Smart Home (SH). The motivation for
PAR minimization in residential sector is due to two reasons: (i) industrial and
commercial users are reluctant to alter their energy utilization pattern: (ii) about
40% of the total energy is consumed by the residential sector. Furthermore, about
65% of electricity consumption can be minimized by SHs and smart buildings
[7]. The main contribution of this paper are following:
Energy Efficient Scheduling of Smart Home 69
• An IAPR is developed to curtail the PAR.

• A comparison analysis of renowned meta-heuristic techniques with the pro-
posed IAPR is performed.
The rest of this paper is organized as follows. Literature review of the existing
schemes is described in Sect. 2. Sections 3 and 4 discusses the problem statement
and proposed system model. Scheduling techniques are presented in Sect. 5. Dis-
cussion and results are presented in Sect. 6 and finally Sect. 7 concludes the
paper.
2 Related Work
For the last few years, many DSM strategies have been developed to optimize
the energy utilization in residential sector with varying objectives.
In [8], a SHEMS is developed to monitor and schedule household electrical
appliances in a conventional house to minimize electricity consumption cost. In
this scheme, if the load on the grid exceeds a predefined limit, the new appli-
ances are either shifted to other time periods or operated using battery power.
Batteries are either charged from the utility in the low price intervals or using
Photovoltaic (PV) cells in day time. Experimental results demonstrate that the
scheme reasonably minimizes electricity consumption cost. However, this work
did not consider the various pricing schemes. Yang et al. [9] present a scheme
using Improved Particle Swarm Optimization (IPSO) for scheduling home appli-
ances. The scheme mainly focuses on minimizing cost. It is evident that elec-
tricity tariff and objective curve have an inverse relationship with each other.
Experimental results show that the desired objective curve is achieved. More-
over, this work aims to achieve power system stability. For this purpose, the
high electricity consumption is reduced in high price intervals. As a result, UC
is compromised.
The authors in [10] propose a Mixed Integer Linear Programming (MILP)
based cost minimization and load equalizing scheme for domestic users. In this
scheme, electricity consumer is considered as a prosumer. Prosumer is the pro-
ducer as well as the consumer of electricity. The work considers that prosumers
have installed PV cells in their homes for electricity generations along with a
connection to a commercial grid in order to fulfill their energy requirements.
In the proposed work, prosumers store energy in batteries only when electricity
generation exceeds the user demand. However, this work ignores exporting sur-
plus energy to maximize their income. The authors in [11] develop a fuzzy based
energy management system. In this work, consumer is involved in changing the
status of appliances from ON-state to OFF-state. The work mainly focuses on
minimizing cost along with PAR. The UC in this scheme is compromised due to
increase in the appliances waiting time.
Israr et al. [12] present an optimization function to maximize the UC. The
proposed scheme aims to retain an intended environment inside a smart home.
Genetic Algorithm (GA) and PSO techniques are used to enhance the UC. For
70 S. Khan et al.
this purpose, the recorded data for a month in an indoor laboratory environment
is analyzed. Experimental results demonstrate that the proposed scheme achieves
desired results. However, this work did not consider electricity consumption cost
and PAR. In [13], a hybrid technique using Harmony Search Algorithm (HSA)
and Enhanced Differential Evaluation (EDE) algorithm is developed. The pro-
posed scheme formulate the demand and supply problem as Multiple Knapsack
Problem (MKP). The work incorporates ESS to maximize the UC and alleviate
PAR. Experimental results demonstrate that the hybrid scheme reasonably min-
imizes the consumer bill as compare to HSA and EDE. However, PAR reduction
using EDE is better as compared to their proposed technique. A chance con-
strained DR scheme is proposed in [14] to schedule electrical appliances in HEMS.
The scheme uses PSO and two-Point Estimation Method (2-PEM). Experimen-
tal results demonstrate that PSO along-with 2-PEM outperforms gradient based
PSO with Latin hyper cube sampling in terms of minimizing computational load.
However, this work, ignores PAR and consumption cost. PAR burdens the utility
that has a direct impact on consumer.
Wen et al. [15] propose HEMS using reinforcement learning technique to
meet the DR of residential buildings. The scheme adds intelligence to the HEMS
via a three step Observe, Learn and Adopt algorithm (OLA). In this technique,
an algorithm learns the electricity consumption pattern of the appliances based
on consumer preferences. Experimental results demonstrate that the proposed
scheme reasonably minimizes the user consumption cost and PAR. However, UC
is compromised due to shifting or delaying the operational time of the electricity
consuming appliances. The work in [16] minimizes PAR in SG via DR strategies.
This work considers the optimization problem as two stage problem. The con-
sumers try to boost their pay-offs and the utility tries to maximize its revenue.
At the consumer end, electricity consumption cost is calculated using an itera-
tive algorithm. Whereas, at the retailer end, the work used Simulated-Annealing
based Price Control Algorithm (SAPCA). SAPCA is used to maximize retailer’s
profit using RTP scheme. Experimental results demonstrate that the proposed
scheme is beneficial for the consumers as well as utility. Furthermore, the pro-
posed work reduces the PAR.
Farrokhifar et al. in [17] develop a scheme for scheduling home appliances to
reduce consumption cost within a building. The scheme considered optimization
problem as the ILP problem. Simulations results demonstrate that the proposed
scheme reduces consumption cost. However, this work ignores the PAR. Khan
et al. in [18] present a priority induced DSM scheme to efficaciously schedule
the home appliances in SHEMS. The main objective of this work is to mini-
mize cost and mitigate rebound peaks. In this work, the authors used renowned
meta-heuristic techniques namely EDE, BPSO, and GA. Furthermore, knapsack
capacity limit is used to enhance the grid reliability. Simulation results shows
that this scheme minimized the consumer bill by 60%. However, this work con-
sidered a limited number of home appliances.
3 Motivation and Problem Statement

The common objectives of SG are cost minimization, optimal load management,
reliability and sustainability of power system and PAR reduction etc. PAR mini-
mization is one of the most challenging task in SG due to the irregular consump-
tion pattern of electricity in residential sector. The work in [15,16] minimized
PAR and energy consumption cost using the meta-heuristic techniques. How-
ever, these scheme compromised UC. In this paper, like the work presented in
[18] the main focus is to minimize the PAR and rebound in order to enhance the
reliability of power system. Although many techniques exists in the literature
for PAR minimization. In this paper, an enhanced meta-heuristic technique is
developed to tackle the PAR optimization problem in SH.
4 System Model
This section discusses the proposed system model. In this paper, the aim is to
shift the execution time of home appliances such that PAR and cost is mini-
mized. A detailed description of our proposed system model is shown in Fig. 1.
To meet the energy demand, it is assumed that the SH is connected to util-
ity using AMI. AMI allows bi-directional communication between the consumer
and utility. Generally AMI comprises of a SM, energy management controller and
data management system. The SM acts as a communication gateway between
the consumer and the utility.
Utility
Smart meter
Fig. 1. System model
4.1 Appliances Categories

Considering Fig. 1, the proposed SHEMS consists of various appliances. Based
on their distinguishing role in PAR and cost minimization these appliances
are divided into three sub-categories namely shiftable and interruptible (ASI ),
shiftable and non-interruptible (ASN ) and regularly operated appliances (AR ).
In our proposed SHEMS, the appliances considered in [19] are used. Table 1 pro-
vides an overview of the home appliances, their power ratings and Length of
Operational Time (LOT).
72 S. Khan et al.
Table 1. Appliances classification
Class Name Power (kWh) LOT (h)

ASI Vacuum cleaner 0.7 6
Electric car 2.5 14
Laptop 0.1 8
Dish washer 1.8 8
Desktop 0.3 6
ASN Washing machine 1 3
Cloth dryer 3 5
AR Cooker top 2.15 2
Cooker oven 5 2
Microwave 1.7 2
Refrigerator 0.3 24
Interior lights 0.84 6
4.2 Problem Formulation

In this paper, the appliances scheduling problem is formalized as an optimization
problem where the objective is to minimize the PAR. The objective of our work
can be formulated as:
minimize P AR s.t. CS < CU (1)
Here CS and CU represents the energy utilization cost in the scheduled and
unscheduled scenario. In order to compute the PAR, Eq. 2 is used.
LP eak
P AR = (2)
LAverage
Here LP eak is the maximum load consumed by the home appliances in SH
for a single time slot and LAverage is the average load consumed in one day. To
compute the per hour and total load consumed by the SH we use Eqs. 3 and 4.
LoadH = ρ(t) ∗ α (3)

24
LoadT = LoadH (t) (4)
t=1
Here LoadH represents the load consumed in one hour, LoadT is the total
load consumed in one day, ρ(t) shows the power consumed by an appliance and
α is the ON and OFF status of an appliance.
5 Scheduling Techniques
This section discusses the existing and proposed scheduling schemes.
5.1 Strawberry Algorithm (SA)

Plants are connected to the earth by means of their roots. It is evident that these
roots cannot move to different places like other animals and birds which migrate
to different places for their survival. However, there exist some grasses and plants
(i.e., Strawberry plant) which propagates by means of runners (stolon). Keeping
in view the vegetative propagation nature of strawberry plants Merrikh-Bayat
in [20] proposed SA. Strawberry plant propagates by means of runners as well
as roots hair. If the strawberry plant is in a location which is not favorable for
its survival, the plant performs exploration. Exploration in strawberry plants
is the process of sending long runners in different directions for its survival. A
runner is a creeping stalk which emerges from leaf axils of a parent or mother
plant. If a runner succeeds in locating a favorable condition for its survival, it
generates additional roots hairs and runners which directly effects the growth
rate of strawberry plant. SA for optimization problem can be modeled using the
following facts.
• All the strawberry plants propagates by means of runners. These runners
arise randomly. Runners explore the global search space.
• All the strawberry plants randomly generate roots and root hairs. It allows
the plant to exploit the local search space.
• If a strawberry plant have easy access to nourishing resources, the plant will
grow faster. Furthermore, that plant will generate more runners and roots.
Whereas, if a plant do not have access to nourishing resources, they have a
high death rate.
5.2 Salp Swarm Algorithm (SSA)

Inspired by the swarming behavior of Salp swarms, Mirjalilia et al. in [21] pro-
posed SSA. Salps belongs to the family of Salpidae. They possess a transparent
barrel-shaped body. Their tissue structure resembles the tissue structure of jelly
fishes. Moreover, their movement behavior also resembles jelly fish i.e., to move
forward, they pumped water through their body as repulsion. The biological
research about Salp swarms is at its early stages. Because Salps are mostly
found in deep oceans. Furthermore, it very difficult to create a favorable living
environment for salps in a laboratory. In deep oceans, Salps form a Salp chain.
The main reason for forming a Salp chain is not yet clear. However, scientists
believe that foraging is one of the many reasons. The chain is divided into two
groups i.e., leader and followers. The leader of the salp chain is at the front
whereas the followers follow their leader directly or indirectly. For modeling the
Salp swarming behavior to optimization problems, it is assumed that there exist
a food source in the search space. The leading Salp targets the food source by
changing its position. Whereas, the followers gradually follow the leading Salp.
5.3 IAPR
It is evident that plant survives by their vegetative propagation nature. For
their purpose, plants send runners and root hair randomly. If the runners or the
74 S. Khan et al.
roots of a plant succeed in locating a nourishing resource, the plant survives,

otherwise the plant dies. Motivated by the vegetative propagative nature of
plants, we proposed a survival based meta-heuristic scheme. In this scheme,
first a pre-defined number of explorers are initialized. With a rich nourishing
resource available in the search space, these explorers move towards that source
independently. It is possible that some of the explorers may not be able to reach
the food source. This might endanger the plant survival. Therefore, to tackle
this problem, we categorize the explorers as best and worst explorers. A best
explorer is the one who succeed in locating the nourishing resource whereas, the
worst explorer is the one who failed to do so. In this scheme to prevent the worst
explorers from endangering the plant survival, the worst explorers are mutated
with best explorers. If the new explorers produced by mutation move towards
the food source, the explorers are updated by replacing the old explorers with
the new ones else if the new explorers do not move in the direction of nourishing
resources, they are discarded. In modeling the proposed IAPR to an optimization
problem, in this paper the food source is termed as the global optimum solution.
However, the global optimal solution to any optimization problem is not known
in advance, therefore, in this scheme, it is assumed that a nourishing source
located first will remain a global solution unless a better nourishing resource is
located. The pseudo code of our propose scheme is given below.
Algorithm 1. Algorithm for IPAR

1: Randomly generate a nourishing resource
2: while stopping criteria not met do
3: Compute c using the current and total iterations
4: for i = 1 to size(explorers) do
5: Movement Probability= rand()
6: if P < M ovement P robability then
7: U pdate explorer position by adding c ;
8: else
9: U pdate explorer position by subtracting c ;
10: end if
11: end for
12: for all explorers do
13: if An explorer lies outside the legal region then
14: P lace it on the boundry ;
15: end if
16: end for
17: Evaluate the f itness of the explorers
18: Classif y the explorers as best and worst ;
19: if N ew explorer moves towards the f ood source then
20: Replace new explorer with old worst explorer ;
21: else
22: Discard new explorer and keep old explorer ;
23: end if
24: end while
25: return operational pattern of home appliances
6 Simulation Results and Discussions

In this section, results of our all the optimization schemes are discussed. The per-
formance of the IAPR is compared with SSA and SA in PAR reduction. Results
are obtained for two types of pricing tariffs namely CPP and RTP. For simula-
tion MATLAB R2017a installed on Intel(R) Core(TM) i5-3380M with 2.90 GHZ
processor is used. Due to the random nature of meta-heuristic optimization tech-
niques, we plot the mean results of ten runs.
4
PAR
0
Unscheduled SA SSA IAPR
Fig. 2. PAR reduction in CPP scheme
4
PAR
0
Fig. 3. PAR reduction in RTP scheme
Figures 2 and 3 represents that IAPR outperforms the well known meta-
heuristic techniques SSA and SA. The proposed IAPR minimized PAR using
CPP by 72%. Whereas, SA and SSA minimized PAR by 51% and 9% respectively.
In RTP, PAR alleviation is 17%, 48% and 68% in SSA, SA and IAPR respectively.
Figures 4 and 5 shows the per hour load consumption pattern of SH. These
figures gives a complete picture of the performance of IAPR using CPP and
RTP. The IAPR balanced the per hour load consumption, as a result PAR in
minimized. The maximum load recorded in a single time slot using CPP and RTP
schemes is 10.54, 7.80, 9.60 and 7.80 kWh for unscheduled, SA, SSA and IAPR
respectively. Here it is worth mentioning that PAR is minimized by balancing
load consumption throughout the day and total load consumed by all algorithms
in both of the pricing scheme was same.
76 S. Khan et al.
12
Per hour load (kWh)

10
4
2
0
1 4 8 12 16 20 24
Time (hours)
Fig. 4. Per hour load using CPP scheme
12
Per hour load (kWh)

10
8
6
2
0
1 4 8 12 16 20 24
Time (hours)
Fig. 5. Per hour load using RTP scheme
Figures 8 and 9 shows the performance in overall cost reduction. It can be

seen from these figures that all the algorithms minimized cost. In CPP tariff,
cost reduction as compare to unscheduled scenario is 34%, 76% and 32% using
SA, SSA and IAPR respectively. Whereas, in RTP scheme cost is minimized by
15%, 40% and 17% using SA, SSA, and IAPR. The proposed IAPR balanced the
hourly load consumption which has a direct impact on the total cost. Figures 6
and 7 shows the per hour cost consumption is SH. These figures reveal that
IAPR successfully minimized cost by load shifting. Furthermore, cost reduction
in IAPR with CPP scheme is higher as compare RTP scheme.
1200
Unscheduled
Per hour cost (cents)
1000 SA
SSA
800 IAPR
600
400
200
0
1 4 8 12 16 20 24
Time (hours)
Fig. 6. Per hour cost using CPP scheme

200 Unscheduled
Per hour cost (cents)

SA
SSA
150 IAPR
100
50
0
1 4 8 12 16 20 24
Time (hours)
Fig. 7. Per hour cost using RTP scheme
5000
Totalcost(cents)
4000
3000
2000
1000
0
Fig. 8. Total cost using CPP scheme
1500
Total cost (cents)
1000
500
0
Fig. 9. Total cost using RTP scheme
The waiting time of all the algorithms using CPP and RTP scheme is shown
in Figs. 10 and 11. The proposed IAPR has a higher waiting time as compare to
the other optimization algorithms. It can be seen from Figs. 2, 3, 4, 5, 6, 7, 8, 9,
10 and 11 that all the algorithms with both the pricing strategies are confronted
with trade-offs. SSA minimized the electricity consumption cost. However, the
performance in PAR minimization is not good and multiple peaks are created
in Off-peak hours. SA minimized the PAR as compare to SSA. However its
performance in waiting time reduction is higher than SSA. Similarity IAPR
minimized PAR, however, it increased the appliances waiting time as compare
to SA and SSA.
78 S. Khan et al.
Waiting time (hours)

6
0
Fig. 10. Waiting time (hours) with CPP scheme
7
6
Waiting time (hours)
5
4
3
2
1
0
Fig. 11. Waiting time (hours) with RTP scheme
7 Conclusion and Future Work
In this paper, an IAPR to minimize the PAR is developed. Performance of the

IAPR is evaluated in the scenario of a SH. This scheme gives the residential
consumers an opportunity to eagerly participate in enhancing the reliability of
power system. To validate the effectiveness of IAPR, the performance in PAR
reduction is compared with the renowned meta-heuristic techniques namely SA
and SSA using the ToUP and CPP scheme. Simulation results confirm that the
IAPR in PAR minimization with both the pricing schemes exceeds SA and SSA.
PAR minimization with the CPP in the IAPR, SA and SSA is 72%, 51% and 9%.
Whereas, in the ToUP scheme PAR is minimized by 17%, 48% and 68% using
SSA, SA and IAPR. In future, we plan to extend this work by incorporating MG
and ESS at consumers’ end to enhance the stability of power grid.
References
1. Nadeem, J., Javaid, S., Abdul, W., Ahmed, I., Almogren, A., Alamri, A., Niaz,
I.A.: A hybrid genetic wind driven heuristic optimization algorithm for demand
side management in smart grid. Energies 10(3), 319 (2017)
2. Farhangi, H.: The path of the smart grid. IEEE Power Energy Mag. 8(1) (2010)
3. Hussain, B., Javaid, N., Hasan, Q., Javaid, S., Khan, A., Malik, S.: An inventive
method for eco-efficient operation of home energy management systems. Energies
11(11), 3091 (2018)
4. Khalid, A., Javaid, N., Mateen, A., Ilahi, M., Saba, T., Rehman, A.: Enhanced
time-of-use electricity price rate using game theory. Electronics 8(1), 48 (2019)
5. Economics, Frontier, and Sustainability First: Demand side response in the domes-
tic sector-a literature review of major trials. Final report, London, August 2012
6. Shareef, H., Ahmed, M.S., Mohamed, A., Al Hassan, E.: Review on home energy
management system considering demand responses, smart technologies, and intel-
ligent controllers. IEEE Access 6, 24498–24509 (2018)
7. Lior, N.: Sustainable energy development: the present (2009) situation and possible
paths to the future. Energy 35(10), 3976–3994 (2010)
8. Shakeri, M., Shayestegan, M., Reza, S.S., Yahya, I., Bais, B., Akhtaruzzaman, M.,
Sopian, K., Amin, N.: Implementation of a novel home energy management sys-
tem (HEMS) architecture with solar photovoltaic system as supplementary source.
Renewable Energy 125, 108–120 (2018)
9. Yang, H.-T., Yang, C.-T., Tsai, C.-C., Chen, G.-J., Chen, S.-Y.: Improved PSO
based home energy management systems integrated with demand response in a
smart grid. In: 2015 IEEE Congress on Evolutionary Computation (CEC), pp.
275–282. IEEE (2015)
10. Van Der Stelt, S., AlSkaif, T., van Sark, W.: Techno-economic analysis of household
and community energy storage for residential prosumers with smart appliances.
Appl. Energy 209, 266–276 (2018)
11. Collotta, M., Pau, G.: A novel energy management approach for smart homes using
bluetooth low energy. IEEE J. Sel. Areas Commun. 33(12), 2988–2996 (2015)
12. Ullah, I., Kim, D.H.: An improved optimization function for maximizing user com-
fort with minimum energy consumption in smart homes. Energies 10(11), 1818
(2017)
13. Kazmi, S., Javaid, N., Mughal, M.J., Akbar, M., Ahmed, S.H., Alrajeh, N.: Towards
optimization of metaheuristic algorithms for IoT enabled smart homes targeting
balanced demand and supply of energy. IEEE Access (2017)
14. Huang, Y., Wang, L., Guo, W., Kang, Q., Qidi, W.: Chance constrained optimiza-
tion in a home energy management system. IEEE Trans. Smart Grid 9(1), 252–260
(2018)
15. Wen, Z., O’Neill, D., Maei, H.: Optimal demand response using device-based rein-
forcement learning. IEEE Trans. Smart Grid 6(5), 2312–2324 (2015)
16. Qian, L.P., Zhang, Y.J.A., Huang, J., Wu, Y.: Demand response management via
real-time electricity price control in smart grids. IEEE J. Sel. Areas Commun.
31(7), 1268–1280 (2013)
17. Farrokhifar, M., Momayyezi, F., Sadoogi, N., Safari, A.: Real-time based approach
for intelligent building energy management using dynamic price policies. Sustain.
Cities Soc. 37, 85–92 (2018)
18. Khan, A., Javaid, N., Ahmad, A., Akbar, M., Khan, Z.A., Ilahi, M.: A priority-
induced demand side management system to mitigate rebound peaks using multiple
knapsack. J. Ambient Intell. Humanized Comput., 1–24 (2018)
19. Pooranian, Z., Abawajy, J.H., Conti, M.: Scheduling distributed energy resource
operation and daily power consumption for a smart building to optimize economic
and environmental parameters. Energies 11(6), 1348 (2018)
20. Merrikh-Bayat, F.: A numerical optimization algorithm inspired by the strawberry
plant. arXiv preprint arXiv:1407.7399 (2014)
21. Mirjalili, S., Gandomi, A.H., Mirjalili, S.Z., Saremi, S., Faris, H., Mirjalili, S.M.:
Salp Swarm Algorithm: a bio-inspired optimizer for engineering design problems.
Adv. Eng. Softw. 114, 163–191 (2017)
Minimizing Daily Cost and Maximizing
User Comfort Using a New Metaheuristic
Technique
Raza Abid Abbasi1 , Nadeem Javaid1(B) , Sajjad Khan1 , Shujat ur Rehman2 ,

Amanullah2 , Rana Muhammad Asif3 , and Waleed Ahmad1
1
COMSATS Institute of Information Technology, Islamabad 44000, Pakistan
2
Quaid-i-Azam University, Islamabad 44000, Pakistan
3
NCBA&E, Multan, Pakistan
Abstract. A home energy management system intended to improve the

energy consumption pattern in a smart home is proposed in this research.
The objective of this work is to handle the load need in an adequate man-
ner such that, electrical energy cost and waiting time is minimized where
Peak to Average Ratio (PAR) is maintained through coordination among
appliances. The proposed scheme performance is assessed for PAR, user
comfort and cost. This work assess the behavior of advised plan for real-
time pricing and critical peak pricing schemes.
Keywords: Smart Grid · DSM · HEMS · BCS
1 Introduction
In recent decades an exponential growth in the energy consumption is observed,
therefore, the gap between demand and generation is increasing. Because of the
rise in the energy need, the conventional grid is facing a number of problems
like reliability, maintenance and sustainability. Residential consumer are grow-
ing rapidly at a pace of 0.3% annually, and has large share of 27% among others
in electricity consumption [1]. In past most of the power generation was achieved
using fossil fuels to fulfill the gap between energy demand and generation. How-
ever in recent years scientists have worked on exploring new means of energy
generation, i.e., renewable and sustainable energy resources [2]. The addition of
Renewable Energy Sources (RESs) resulted in an increase in the power system
complexity, where existing grids were not able to manage that. Smart Grid (SG)
is introduced to handle the above mentioned challenges. A SG is equipped with
Information and Communication Technology (ICT), that enabled it to incorpo-
rate RESs and manage the stability and reliability of power systems. The most
important feature of SG is the control of energy production, transmission and
distribution through advanced ICTs.
https://doi.org/10.1007/978-3-030-15035-8_8
HMS Using Hybrid Meta-heuristic Optimization Technique 81
The state of the art techniques for DSM and SG are discussed in Sect. 2.
Different problems related to the DR and DSM are reflected along with their
consequences in Sect. 3. Section 4 explains the suggested scheme for discussed
problemettes. Section 5 highlights formerly employed techniques, Strawberry
Algorithm (SBA), Earthworm optimization Algorithm (EWA), and schemed
algorithm. Simulations outcomes are depicted and shown in Sect. 6. At last,
Sect. 7 draws conclusion of the work done.
2 Related Work
SG enabled the utility to manage the DR of customers to minimize the overall

load at power generating sources.
In [6] Huang et al. proposed to enhance the control of appliances in a HEMS,
considering the uncertain electricity loads and cost. A new scheme, involving,
Gradient-based Particle Swarm Optimization (GPSO) for solving the global
optimum solution and Two Point Estimation Method (2PEM) for handling the
indeterminate behavior of must load runs was introduced. Thus, the presented
GPSO-2PEM can minimize processing cost and can be easily applied in an
embedded device having limited resources. Hansen et al. proposed HEMS scheme
for reducing the household electricity bill in [7] using RTP pricing scheme. They
introduced non-myopic Partially Observable Markov Decision Process (POMDP)
technique for reducing the peak demand which ultimately reduces the overall
electricity cost.
In [8] authors proposed a new HEMS which aims at minimizing the next day
energy cost and disturbance to the user with RTP and household photo-voltaic
penetrations. In proposed system, first user defined electricity usage restrictions
are set. Then, an optimal scheduling model is proposed based on the predicted
solar output and energy prices. It will support the decision making for the RES
operations. They used advanced adaptive thermal comfort model for sensing the
users indoor thermal comfort degree which supports in scheduling the heating,
ventilating and air conditioning systems. Where, user disturbance value met-
ric is suggested to understand the psychological disturbance of an appliance
schedule on the users preference. Luo et al., worked on optimize scheduling of
Distributed Residential Energy Resources (DRERs) in [9]. They worked for a SH
having varying electricity pricing and high RES penetrations. They used Monte
Carlo sampling technique for handling the uncertainties of solar power output.
Users preferences are also taken into account with different levels of priority,
i.e., normal and high. Considering all this, an optimal DRER scheduling model
is proposed to minimize the electricity bill. Authors used natural aggregation
algorithm to solve the proposed model.
Elghitani et al., proposed a methodology for residential demand aggrega-
tion, based on a multi-class queuing system in [10]. Authors used this model for
reducing the cost of the energy power consumption below day-ahead predicted
pricing. Proposed scheme achieves a cost that is close to the best solution. In [11]
Mosaddek Hossain et al., proposed a Real-time Decentralized DSM (RDCDSM)
82 R. A. Abbasi et al.
to adjust the real-time residential load to follow a preplanned day-ahead energy

generation by the microgrid. based on predicted customers aggregate load. A
deviation from the predicted demand at the time of consumption is assumed
to result in additional cost or penalty. Wu et al. in [12] considered the energy
storage uncertainties due to intermittent renewable energy supplies and energy
storage opportunity. Authors proposed a new stochastic dynamic programming
model for the optimal energy management of a smart home with Plug-in Electric
Vehicle (PEV) energy storage. They worked on minimizing the energy cost while
satisfying the user energy demand and PEV storage requirements.
In [13] Moon et al., used 2-stage forecasting scheme for prediction of opti-
mal operation of power system in educational buildings using Short Term Load
Forecasting (STLF) model. Authors collected load data of a university campus
for last five year. Where, using the moving average method found the electric
load pattern of day of a week. Random forest method is being used with time
series cross-validation for forecasting the daily electric load. Keles et al. in [14]
proposed a new scheme that uses Artificial Neural Network (ANN) for predicting
the electricity prices. In a ANN forecasting scheme the output performance is
dependent on the parameter sets, therefore, selection and preparation of funda-
mental data that has prominent effect on the electricity prices is in more focus.
3 Problem Statement
From the time of its evolution, SG has a number of complications as well, i.e.,
energy transferring, equality, surveillance and secrecy at DSM level [5]. The
uneven use of electricity on user side rises the burden on the power distribution
and power generating system. Due to irregular usage, energy required in some
hours (peak hours) increases exponentially. Such high demand results in using
extra generation units to provide the required energy, which subsequently results
increase in cost. Therefore, DSM needs a proficient scheme that can distribute
the load between peak hours and off peak hours evenly.
Although, authors have already done a lot of work on DSM for reducing cost,
waiting time and PAR. Authors in [6,10,17,18] worked on DR optimization con-
sidering uncertainties of must run load, aimed at reducing the upper bound of
energy consumption cost, however PAR was not considered. In [7,8,11] authors
worked on reducing the peak demand through scheduling techniques for appli-
ances incorporating the RES, however user comfort was not considered.
4 Proposed System Model

For efficient usage of electricity, DSM is the base for HEMS. Smart appliances in
a SH are connected to one another and Energy Management Controller (EMC)
through Home Area Network (HAN). Energy usage data from smart appliances
is sent to EMC and then EMC process that data for making important deci-
sions. Utility get the information sent from SM and executes required processing
accordingly. Utility propagates the pricing signal and the consumer required
energy to SM. SM sends the information coming through the utility with the
EMC. Now, EMC possesses the data collected from appliances and the cost infor-
mation received from utility. Information collected from utility and appliances
in used by EMC for optimum appliances scheduling. EMC uses the Strawberry
Algorithm (SBA), Earthworm Optimization Algorithm (EWA) and proposed
scheme for scheduling. The goal of the scheduling is to reduce PAR, maximize
user satisfaction through minimizing waiting time and reducing cost. The dis-
cussed system model is displayed in Fig. 1.
In our proposed model, we have a house, using 11 appliances and those appli-
ances are distributed among different classes. We classify appliances in three
main classes, i.e., Fixed, shift-able and interruptible appliances. Fixed appli-
ances require to be ON according to the specified schedule. Those appliances
that are optional to scheduling but the can not be interrupted while they are
working are classified as Shiftable appliances. Furthermore, there is a condition
for washing dryer that it can start its operation when washing machine has com-
pleted its operation. Appliances those are schedulable as well as interruptible are
placed in Interruptible appliances category. Table 1, displays the classified smart
appliances as well as their LoT and power rating in respective columns. The pro-
posed schemes objective is to: minize PAR, maximize user comfort by reducing
waiting time and reducing total cost.
Table 1. Control parameters
Classes Appliances LoT (h) PR (kWh)

Fixed appliances Light 12 0.1
Coffee maker 4 0.5
Oven 9 3
Blender 4 1.2
Shiftable appliances Washing machine 5 0.5
Cloth dryer 4 4
Dish washer 4 1.5
Interruptible appliances Water heater 12 1.1
Space heater 12 1.5
Iron 6 1.1
Vacuum cleaner 5 0.5
5 Metaheuristic Optimization Algorithms

Meta-heuristic algorithms are structured to provide solution to almost any type
of problem. Therefore, we will use metaheuristic techniques for finding the solu-
tion to scheduling home appliances problem. In this study we are dedicated to
total energy consumption cost minimization and user comfort maximization.

Now we will discuss the selected metaheuristic algorithms.
5.1 SBA
SBA is a nature inspired numerical optimization algorithm. This algorithm is
derived from the strawberry plant behavior for searching the resources and is
being used for solving complex computing problems. Strawberry plant uses both
runners and root for searching resources like water and minerals. Where, these
runners being used for global search and roots being used for local search. Runner
and roots are randomly initialized at the beginning. While algorithm running,
when these runners and roots finds new resources then the reproduction starts
which results in more runners and roots. The locations of roots and runners at
a specific iteration is calculated using Eq. 4. Where, Xroot (i) and Xrunner (i) are
matrices containing locations of roots and runners respectively. Equations 5 and
6 are used for optimization and is the mathematical representation of strawberry
algorithm.
5.2 EWA
Earthworm executes numerous optimization runs during the reproduction pro-
cess. Reproduction process of the earthworm follows following steps.
• Every earthworm is capable of participating in regeneration process, it can
reproduce at most two earthworms.
• The singular generated earthworm contains the complete genetic behavior of
its parent that is of same length as parent.
• The single fit earthworm supports a straight next generation and can not be
changed by operators. It guarantees that earthworm can not end with the
increment in generations.
5.3 Proposed
Our proposed scheme is based on the behavior of plant roots and earthworm.
Plant root search resources, i.e., minerals and water and grows in the direction of
best possible route to the resource. Earthworm performs numerous runs during
the reproduction process. Every earthworm has the ability to reproduce and has
same capability of reproduction like parent earthworm. In our proposed scheme
we consider plants as solution to the problem, plant roots for finding the global
optimum solution and earthworm capability for local optimum solution. The
resource found quality is mapped as fitness function in our proposed scheme.
The steps followed in our proposed scheme are listed below.
Algorithm 1. Proposed Scheme Algorithm for Appliance Scheduling.

Require: Input: [Initialized the population]
while iteration < M axN umbererIteration do

for i ← 1 to size(population) do
Generate offspring through Reproduction 1
Generate offspring through Reproduction 2
end for
for j ← 1 to size(P rimes) do
Perform Crossover
end for
Evaluation of fitness
for i ← 1 to size(population) do
Select random location by chance as mother of the next iteration
end for
Apply mutate
Local best solution is extracted for iteration j
end while
Global best solution is achieved
Appliances are scheduled
6 Simulation Results and Discussion

We evaluated the performance of our proposed technique for reducing the cost.
The results obtained from extensive simulation are discussed here in this section.
We compared results obtained from simulations with other techniques, i.e.,
Earthworm optimization Algorithm (EWA) and Strawberry Algorithm (SBA).
Critical Peak Pricing (CPP) and Real Time Pricing (RTP) pricing signals are
used to examine the usage pattern of the electricity by the users. We consid-
ered single home with eleven appliances, i.e., D = 11 for simulation purpose.
We classified these appliances in three different classes, i.e., Fixed Appliances
(FA), Shiftable Appliances (SA) and Interruptible Appliances (IA). These appli-
ances are selected due to their frequent use in winter. Light, coffee maker, oven
and blender are included in FA. FA can neither be shifted nor interrupted.
Washing machine, cloth dryer and dish washer are included in SA. These appli-
ances can not be interrupted once they have started their working, however these
appliances can be shifted. Water heater, space heater, iron and vacuum cleaner
are included in IA. This group of appliances can be shifted as well as interrupted.
CPP and RTP are used for load, cost and PAR calculation.
6.1 Pricing Tariff

As mentioned earlier, we will evaluate our proposed scheme using CPP and RTP
pricing signals. Here we will discuss CPP and RTP pricing tariff.
6.1.1 CPP Tariff

During hot summer weekdays utilities face emergency conditions. Due to critical
condition during a specific time period utility impose high price rate. The energy
cost through this interval are commonly above average. CPP has two different
types. In the first variant specific instance of time and amount of peak prices is
already known. In the second variant, electricity prices changes with the use of
electricity. If electricity use is increased electricity price also increases to reduce
the load on the utility. Maximum fifteen time these critical hours are allowed in
a season.
6.1.2 RTP Tariff

The RTP pricing scheme is also called dynamic price rate due to its varying
nature. It depends on amount of electricity used per hour. It feeds details about
actual cost of power at a specific time to the user. It empowers the user to
acclimate their working hours from on peak hours to off peak hours, which
results in saving. It is aided using SM which allows the mutual exchange of
information between utility and consumer. Utilities calculate the bill in RTP as
sum of two different elements.
(a) Base cost is computed based on Customer Baseline Load (CBL). It is a
standard defined tariff.
(b) Per hour usage cost is put in based on the time of use of electricity. This is
the basic difference between CBL and the actual usage.
Figures 4 and 5 are showing the CPP and RTP power consumption pattern for
unscheduled and scheduled schemes, i.e., SBA, EWA and proposed respectively.
The proposed scheme is scheduling the energy in an efficient way in comparison
to alternatives. However, proposed scheme load consumption in last hour is high
which is due to reducing load consumption during peak hours (Fig. 2).
Figures 6 and 7 are showing the unscheduled, SBA, EWA and proposed scheme
hourly cost respectively for CPP and RTP pricing schemes respectively. It is
evident from figures that proposed scheme is performing better then other schedul-
ing schemes while transferring the high energy demand to off peak hours. Further-
more, we can see that even SBA and EWA are performing better than unscheduled
(Fig. 3).
Fig. 2. CPP per hour load
Fig. 3. RTP per hour load
The PAR for unscheduled, SBA, EWA and proposed scheme using CPP
and RTP pricing schemes is displayed in Figs. 8 and 9 respectively. In case of
CPP pricing scheme PAR is reduced by 58.77%, 62.29% and 58.15% using SBA,
EWA and proposed scheme. Here, EWA outperformed other schemes. In case of
RTP pricing scheme PAR is reduced by 54.99%, 54.86% and 51.42% using SBA,
EWA and proposed scheme. Here, SBA outperformed other schemes. It is worth
mentioning that all schemes performed better than unscheduled.
Figures 10 and 11 are displaying the waiting time for unscheduled, SBA, EWA
and proposed scheme using CPP and RTP pricing schemes respectively. In case
of CPP pricing scheme waiting time is reduced to 23.37%, 25.41% and 18.77% for
SBA, EWA and proposed scheme. Here, proposed scheme outperformed other
schemes. In case of RTP pricing scheme waiting time is reduced to 24.16%,
24.28% and 18.13% using SBA, EWA and proposed scheme. Here, proposed
scheme outperformed other schemes. It is also noticed that all schemes performed
better than unscheduled.
Fig. 4. CPP hourly cost
Fig. 5. RTP hourly cost
Fig. 6. CPP peak to average ratio

Fig. 7. RTP peak to average ratio
Fig. 8. CPP waiting time
Figures 12 and 13 are displaying the total cost for unscheduled, SBA, EWA
and proposed scheme using CPP and RTP pricing schemes respectively. In case
of CPP pricing scheme cost is reduced by 18.75%, 30.65% and 26.99% using
SBA, EWA and proposed scheme. Here, EWA outperformed other schemes. In
case of RTP pricing scheme cost is reduced by 13.99%, 15.25% and 27.05% using
SBA, EWA and proposed scheme. Here, our proposed scheme outperformed other
schemes.
The proposed scheme minimizes the waiting time, however, it compromises
for PAR and cost in case of CPP pricing signal. Where, proposed scheme mini-
mizes the waiting time and cost, however, it compromises PAR in case of RTP
pricing signal.
Fig. 9. RTP waiting time
Fig. 10. CPP total cost
Fig. 11. RTP total cost

7 Conclusion
DSM is focused to manage the appliances on the user end for efficient use of
energy. Load shifting technique of DSM is used in proposed HEMS to attain
the desired results. A single house consisting of various appliances is considered
for scheduling. Each appliance of house is scheduled using three heuristic tech-
niques; SBA, EWA, and proposed scheme. The proposed scheme helps to find the
most optimal schedule of each house appliance considering the constraints. The
performance of applied techniques is assessed with respect to basic parameters:
PAR, user comfort and energy consumption cost. The cost of energy consump-
tion is calculated using RTP and CPP pricing schemes. Control parameters and
categorization of appliances is kept same, however, power rating of electrical
devices and their types is kept same for a fair comparison. Results are depicting
that waiting time is reduced to 18.77% in case of CPP and 18.13% for RTP
pricing scheme which is best with respect to SBA and EWA. Cost is reduced by
27.05% in case of RTP which is best in comparison to the other two techniques.
However, there is a trade-off, as we cannot achieve the best values for all three
performance parameters simultaneously.
References
1. U.S. Energy Information Administration, Annual Energy Outlook 2018, p. 121,
122, Residential and commercial energy consumption grows gradually from 2017
to 2050. https://www.eia.gov/outlooks/aeo/pdf/AEO2018.pdf. Accessed October
2018
2. Luo, F., Ranzi, G., Wan, C., Xu, Z., Dong, Z.Y.: A multi-stage home energy
management system with residential photovoltaic penetration. IEEE Trans. Ind.
Inf., 11 (2018)
3. Strbac, G.: Demand side management: benefits and challenges. Energy Policy 36,
4419–4426 (2008). Energies 11, 1426 (2018)
4. Gellings, C.W.: The concept of demand-side management for electric utilities. Proc.
IEEE 73(10), 1468–1470 (1985)
5. Baharlouei, Z., Hashemi, M.: Demand side management challenges in smart grid:
a review. In: Proceedings Smart Grid Conference (SGC), pp. 96–101, December
2013
6. Huang, Y., Wang, L., Guo, W., Kang, Q., Wu, Q.: Chance constrained optimization
in a home energy management system. IEEE Trans. Smart Grid 9(1), 252–260
(2018)
7. Hansen, T.M., Chong, E.K.P., Suryanarayanan, S., Maciejewski, A.A., Siegel, H.J.:
A partially observable markov decision process approach to residential home energy
management. IEEE Trans. Smart Grid 9(2), 1271–1281 (2018)
8. Luo, F., Ranzi, G., Kong, W., Dong, Z.Y., Wang, F.: Coordinated residential energy
resource scheduling with vehicle-to-home and high photovoltaic penetrations. IET
Renew. Power Gener. 12(6), 625–632 (2018)
9. Luo, F., Dong, Z.Y., Xu, Z., Kong, W., Wang, F.: Distributed residential energy
resource scheduling with renewable uncertainties. IET Gener. Transm. Distrib.
12(11), 2770–2777 (2018)
10. Elghitani, F., Zhuang, W.: Aggregating a large number of residential appliances for
demand response applications. IEEE Trans. Smart Grid 9(5), 5092–5100 (2018)
11. Tushar, M.H.K., Zeineddine, A.W., Assi, C.: Demand-side management by regu-
lating charging and discharging of the EV, ESS, and utilizing renewable energy.
IEEE Trans. Ind. Inf. 14(1), 117–126 (2018)
12. Wu, X., Hu, X., Yin, X., Moura, S.J.: Stochastic optimal energy management of
smart home with PEV energy storage. IEEE Trans. Smart Grid 9(3), 2065–2075
(2018)
13. Moon, J., Kim, K.-H., Kim, Y., Hwang, E.: A short-term electric load forecasting
scheme using 2-stage predictive analytics. In: 2018 IEEE International Conference
on Big Data and Smart Computing (2018)
14. Keles, D., Scelle, J., Paraschiv, F.: Wolf Fichtner Extended forecast methods for
day-ahead electricity spot prices applying artificial neural networks. Appl. Energy
162, 218–230 (2016)
15. Ghofrani, M., Hassanzadeh, M., Etezadi-Amoli, M., Fadali, M.: Smart meter based
short-term load forecasting for residential customers. In: North American Power
Symposium (NAPS), pp. 1–5, August 2011
16. Huang, H., Cai, Y., Xu, H., Yu, H.: A multi-agent minority-game- based demand-
response management of smart buildings towards peak load reduction. IEEE Trans.
Comput.-Aided Des. Integr. Circuits Syst. 36(4), 573–585 (2017)
17. Ruelens, F., Claessens, B.J., Vandael, S., Schutter, B.D., Babuška, R., Belmans,
R.: Residential demand response of thermostatically controlled loads using batch
reinforcement learning. IEEE Trans. Smart Grid 8(5), 2149–2159 (2017)
18. Pourmousavi, S.A., Patrick, S.N., Nehrir, M.H.: Real-time demand response
through aggregate electric water heaters for load shifting and balancing wind gen-
eration. IEEE Trans. Smart Grid 5(2), 769–778 (2014)
19. Jindal, A., Dua, A., Kaur, K., Singh, M., Kumar, N., Mishra, S.: Decision tree
and SVM-based data analytics for theft detection in smart grid. IEEE Trans. Ind.
Informat. 12(3), 1005–1016 (2016)
20. Mirowski, P., Chen, S., Ho, T.K., Yu, C.-N.: Demand forecasting in smart grids.
Bell Labs Tech. J. 18(4), 135–158 (2014)
21. Wijaya, T.K., Vasirani, M., Humeau, S., Aberer, K.: Cluster-based aggregate fore-
casting for residential electricity demand using smart meter data. In: IEEE Inter-
national Conference on Big Data (Big Data)
22. Yang, J., Liu, J., Fang, Z., Liu, W.: Electricity scheduling strategy for home energy
management system with renewable energy and battery storage: a case study. IET
Renew. Power Gener. 12(6), 639–648 (2018)
23. Merrikh-Bayat, F.: A numerical optimization algorithm inspired by the strawberry
plant, arXiv preprint arXiv:1407.7399 (2014)
24. Waterloo North Hydro Inc. Residential Charges and Rates. Accessed 20 May 2016.
https://www.wnhydro.com/en/yourhome/time-of-use-rates.asp
Efficiency Analysis of TFHE Fully
Homomorphic Encryption Software Library
Based on GPU
Hai-bin Yang, Wu-jun Yao(&), Wen-chao Liu, and Bin Wei
School of Cryptographic Engineering, Engineering University of the Chinese

People’s Armed Police, Xi’an 710086, Shaanxi, China
sensornet@163.com
Abstract. There are a large number of independent matrix and vector opera-
tions in the lattice-based homomorphic encryption. These operations are suitable
for GPU, which can greatly improve the efficiency of homomorphic operations.
In this paper, we analyze the structure of the homomorphic encryption algorithm
and verify the reliability of the homomorphic encryption software library, debug
and analyze the fully homomorphic encryption software library TFHE and its
corresponding GPU version cuFHE, and then compare their efficiency. The
experimental results show that the GPU version TFHE is 4.5 times faster than
the CPU version TFHE, so the GPU can greatly improve the homomorphic
running speed of the homomorphic encryption scheme.
1 Introduction
The FHE (fully homomorphic encryption) algorithm can solve the problem of data
security and privacy in cloud computing [1]. In cloud computing, the data owner
uploads the data to the server for processing, but the server does not consider the data
to be trusted, so the data needs to be encrypted before uploading. Data encrypted by
traditional data encryption will lose its original characteristics, causing most data
processing methods to fail [2]. FHE can directly operate on ciphertext, and the result of
the operation is equal to the result of doing the same operation on plaintext [3].
In 2009, Gentry proposed the first FHE scheme [4]. In this scheme, Gentry pro-
posed a construction idea: Bootstrapping (running its own decryption circuit with the
homomorphic scheme) + Squashing (squashing the depth of the decryption circuit),
which is also known as the Gentry blueprint.
In order to apply the homomorphic algorithm to the actual situation and improve the
efficiency of the scheme, the researchers continue to improve the bootstrap process [5].
At CRYPTO2009, Gentry, Sahai and Waters proposed a GSW [6] FHE scheme based
on approximate eigenvectors, which uses Cipher Flatten technology to construct a rel-
atively simple homomorphic encryption scheme based on LWE. At ASIACRYPT2016,
Chillotti uses the feature that the product is not greater than the multiplier on the multi-
plication of T = (0,1], constructs a more efficient homomorphism scheme based on the
variant TGSW [7]. It replaces the product of the original TGSW ciphertext (matrix) and
the TGSW ciphertext (matrix) by the external product of TGSW ciphertext (matrix) and

https://doi.org/10.1007/978-3-030-15035-8_9
94 H. Yang et al.
TLWE ciphertext (vector), and constructs the addition on polynomial exponents more
efficiently, which reduce the time of the bootstrapping process to 0.1 s. At ChinaCrypt
2017, Zhou [8] optimized the accumulating operation in the bootstrapping process, which
increased the speed of the single-bit bootstrapping process to 10 ms/gate, and it is the
most efficient single-bit bootstrap FHE scheme.
Recently, GPUs have been widely used in general-purpose computing. The CUDA
enables GPU parallel computing to have better programmability, and programming is
more efficient and flexible. This advantage is also used to further improve the efficiency
of FHE. In 2009, Wang et al. implemented megabit homomorphic multiplication by
GPU parallel computing of FFT and Barrett modular reduction algorithm [9, 10]. In
2015, Wei et al. implemented a GPU-accelerated NTRU scheme by modulo multi-
plication in the remaining space, and processed the large-coefficient polynomial by the
NTT and CRT methods, thereby realizing the GPU optimization of the LTV hierar-
chical homomorphic encryption scheme [11, 12].
TFHE is currently a more efficient single-bit FHE algorithm. In order to reduce the
noise, the scheme reduces the range of the plaintext to (0,1] during the bootstrap
process, and decomposes the ciphertext. In this paper, we debug the TFHE library and
its corresponding GPU accelerated version cuFHE, test its operation mode, and
dynamically compare and analyze their performance. Our goal is to verify the relevant
properties of FHE and analyze the related properties of the algorithm from the per-
spective of program implementation, providing a practical basis for GPU accelerated
homomorphic encryption research.
2 Preliminary
2.1 Algorithm
Let the security parameter to be k, set B , f0; 1g, real number ring with a modulus of 1
T , R=Z 2 ½0; 1Þ, polynomial ring R , Z½X=ðX N þ 1Þ, polynomial ring with a mod-
ulus of 1 TN ½X , R½X=ðX N þ 1Þ, where the 2N-th cyclotomic polynomial
Q
U2N ðXÞ ¼ ðX xi2N Þ ¼ X N þ 1, Mp;q ðEÞ is a matrix size of p q that E
i2Z2N
encrypted.
Bootstrapping: h 2 Mðk þ 1Þl;k þ 1 ðTN ½XÞ, Dech;b;e is the decomposition function, s 2
Bn ; s00 2 BN ½Xk ; a, c is the noise parameter, BK ¼ BKs!s00 ;a is the bootstrapping key,
Coefficient matrix is s0 ¼ KeyExtractðs00 Þ 2 BkN , exchange key is KS ¼ KSs0 !s;c;t ,
LWE ciphertext ða;bÞ 2 LWEs;g ðlÞ, where l 2 T, two related messages l0 and l1 ,
output LWEs ðl0 Þ after bootstrapping. If jus ða; bÞj 14 d, l0 ¼ l0 ; else if
jus ða; bÞj 14 þ d, l0 ¼ l1 , where d is the accumulated deviation, the worst case is n4N
þ1
,
an the best case is 0. If #BK ¼ VarðErrðBKi ÞÞ ¼ p a , VKS ¼ VarðErrðKSi ÞÞ ¼ p c , v
2 2 2 2
is the output of Bootstrapping program, then kErrðvÞk1 2nðk þ 1ÞlbNa þ kNtc þ

nð1 þ kNÞe þ kN2ðt þ 1Þ , kErrðvÞk 2Nnðk þ 1Þlb2 #BK þ kNtVKS þ nð1 þ kNÞe2 þ
kN22ðt þ 1Þ .
Efficiency Analysis of TFHE Fully Homomorphic Encryption Software Library 95
2.2 GPU Parallel Computing

GPU (Graphics processing unit) uses the single instruction multiple threads (SIMT)
execution model, the control unit dispatches instructions to each pipeline. Each pro-
cessor contains multiple sub-processors, this architecture provides a suitable hardware
platform for high-performance computing, as shown in Fig. 1.
TPC TPC TPC TPC ... TPC TPC TPC TPC
TPC: Texture Processor Cluster SM: Streaming Multiprocessor

Instruction L1 Data L1
SM Instruction Fetch/Dispatch
Shared Memory
TEX SP SP
SP SP
SFU SFU
SM SP SP
SP SP
SP: streaming processor
Fig. 1. GPU hardware architecture
The GPU can be used to speed up the FHE bootstrapping process because it
contains thousands of CUDA cores, which allows the graphics card to compute high-
dimensional matrices in parallel. In SIMT mode, the next thread starts the I/O operation
when the previous thread performs the operation. It is similar to the instruction pipe-
line. When the number of threads is enough, the large amount of overhead caused by
I/O can be blocked.
Thread management is an important part of CUDA programming. In CUDA, a
stream represents a GPU operation queue. By default, CUDA creates a default stream
that performs a series of operations in sequence. Figure 2 shows a typical GPU thread
view.
Grid
Thread Block Thread Block Thread Block

N-1 N N+1
Thread Thread Thread Thread Thread Thread Thread Thread Thread

N-1 N-1 N-1 N-1 N-1 N-1 N-1 N-1 N-1
Fig. 2. CUDA thread view

96 H. Yang et al.
Memory management is another important factor affecting the efficiency of GPU

programming. Usually, the parallel computing time is much less than the time of
memory reading and writing and swapping. The GPU includes global memory, local
memory, shared memory, register memory, constant memory, and texture memory.
3 TFHE Library Debug and Analysis

3.1 TFHE Library
The TFHE library is a fully homomorphic encryption library based on C++ language,
which can realize the TFHEW proposed by Ilaria Chillottti et al. [12]. The library
encapsulates the homomorphic rounding function constructed by polynomial multi-
plication, the constant term of the homomorphic truncated plaintext polynomial and the
homomorphic addition on the polynomial order, reducing the time of the bootstrap
process to 52 ms, and realizing Efficient homomorphic constant multiplication of
TGSW ciphertext with plaintext space {0,1} and constant, reduce the bootstrap key
from 1G bytes to 23M bytes.
3.2 TFHE Library Installation and Debugging

Our experiment uses the following software and hardware environment: Ubuntu 16.04
OS, Inter(R) Core(TM) i7-7700HQ CPU, 16 GB memory, NVIDIA GEFORCE GTX
1050Ti graphics card, and CUDA9.2.
Installation and debugging steps are as follows:
Step 1: Install CMAKE. Download the compressed package from the official website,
extract it and enter the folder. Then enter the commands
./bootstrap ! make ! make install.
Step 2: Install fftw3 library. ① Download fftw-3.3.6-pl2.tar.gz from www.fftw.org;
② Unzip tar zxvf fftw-3.3.6-pl2.tar.gz; ③ Configuration parameters. ./configure –
enable-type-prefix –prefix = /usr/local/fftw –with-gcc –disable-fortran –enable-i386-
hacks; ④ Run the compile command make; ⑤ Run the install command make install;
⑥ Run the clear command make clean; ⑦ Reconfigure parameters ./configure –
enable-float –enable-type-prefix –prefix=/usr/local/fftw –with-gcc –disable-fortran –
enable-i386-hack. Repeat steps 5 and 6 until the installation is complete.
Step 3: Install Googletest. Download gtest source code from https://googletest.
googlecode.com. ① Unzip source code package unzip gtest-1.8.0.zip; ② Enter source
file cd gtest-1.8.0; ③ Compile CMakeLists.txt to generate two static libraries libgtest.a
and libgtest_main.a; ④ Run command cp -r include/gtest/ /usr/local/include/, copy the
include/gtest directory to global header file directory.
Step 4: Create and open the build file. Run command mkdir build to create a new
build file, then go to the folder.
Step 5: Compile the tfhe code and configure the relevant parameters. Input command:
cmake ../src -DENABLE_TESTS=on -DENABLE_FFTW=on -DCMAKE_BUILD_
TYPE=debug -DENABLE_NAYUKI_PORTABLE=off -DENABLE_NAYUKI_AVX=off -
DENABLE_SPQLIOS_AVX=off -DENABLE_SPQLIOS_FMA=off
Step 6: Add the search directory to the CMAKELIST.TXT file. Insert into
CMAKELIST.TXT file: set(CMAKE_INCLUDE_PATH “/usr/local/fftw/include/”), set
(CMAKE_LIBRARY_PATH “/usr/local/fftw/lib/”) and include_directories(“/usr/local/
fftw/include/”)
Step 7: Add execution links and commands to the CMAKELIST.TXT file. add_
executable(test-bootstrapping-${FFT_PROCESSOR}testbootstrapping.cpp${TFHE_-
_HEADERS} and target_link_libraries(test-bootstrapping-${FFT_PROCESSOR}
${RUNTIME_LIBS}).
3.3 TFHE Library Debug Process

The debugging process is as Fig. 3 shows:
Begin
Enter folder: build
Command: make
Compile
Enter filefolder: test
Run executable file: ./*
parameters Encryption key

Input
Bootstrap key Exchange key
Plaintext: in_message
Bootstrapping...
Bootstrap Time end_variance

Output
end_phase end_message
Fig. 3. Debugging process
4 cuFHE Library Debugging and Analysis
4.1 cuFHE Library

cuFHE library is an open source GPU accelerated FHE library based on CUDA
platform, it is an instance of GPU-accelerated TFHE scheme [13].
98 H. Yang et al.
4.2 cuFHE Library Installation and Debugging

Run the terminal program in the ‘cufhe’ folder and enter the make command to compile
the default settings.
Step 1: Create buile and bin filefolders.
Step 2: Generate shared library libcufhe_cpu.so
Step 3: Generate executable binaries for test_api_cpu and test_api_gpu in the test/
bin folder, or run make cpu/make gpu to generate executable file.
4.3 cuFHE Library Debug Process

The debugging process is as Fig. 4 shows:
Begin
Enter folder: bin
Command: ./ test_api_gpu
Run executable file: ./*
Generate public key and private key
256 groups of encryption and

decryption tests
GPU initialization
Copy the host data to the device for

processing
256 Nand operations
Output average time required for each gate
Fig. 4. Debugging process
5 Performance Analysis
In order to analyze the performance of the program, we constantly adjust the parameters,
compare the experimental results and analyze the reasons. The ciphertext of the fully
homomorphic scheme will increase the noise after homomorphic operation, so we use
bootstrap technology to reduce the noise of ciphertext to a reasonable range. In the
bootstrap process of TFHE, in order to achieve homomorphic decryption, it is necessary
to complete the homomorphic rounding function round, exponential homomorphic
multiplication and homomorphic addition. In the homomorphic constant multiplication,
the TGSW ciphertext is skillfully constructed, and the most time-consuming is just the
homomorphic accumulation of the TGSW ciphertext, and we conducted an experi-

mental analysis of the variance of the accumulation process. Set the parameters as
N ¼ 1024, n ¼ 500, 1 ¼ 3, bk ¼ 10, then it requires a total of 500 accumulations. The
variance change is shown in Fig. 5, it can be seen that the variance is uniformly
changed.
Fig. 5. Variance during accumulation
In order to verify the correctness of the scheme, we analyzed the error function erf
values of different decomposition parameters. It can be seen from Fig. 6 that the
smaller the component value 1, the larger the error function erf value and the higher the
correctness. For the same 1, when the number of decomposition bits bk < 11, the
correct rate does not change much, but when bk > 11, the correct rate will drop sharply.
Therefore, in order to ensure the correctness, the selected component value 1 cannot be
greater than 11.
Fig. 6. erf trend

100 H. Yang et al.
Accuracy is another measure of FHE scheme. We analyzed the accuracy e of

different decomposition parameters and the results are shown in Fig. 7. It can be seen
that for the same number of decomposition bits, the larger the 1 value, the larger the
precision; for the same component value 1, the precision increases as bk increases.
Fig. 7. Accuracy trend
Under the same hardware conditions, we test TFHE and cuFHE separately. The test
sample is 256 groups. The test time is the average time within the group. The time-
consuming of the key generation algorithm, encryption algorithm and decryption
algorithm, and the average Nand operation time are shown in Table 1. It can be seen
that the encryption and decryption algorithm of cuFHE has been significantly
improved, and the Nand gate operation time on the GPU platform is approximately 4.5
times that on the CPU platform.
Table 1. Time consuming of TFHE v.s. cuFHE

Algorithm KeyGen Enc Nand
TFHE 837.00 ms 0.023 ms 13.00 ms
cuFHE 2758.32 ms 0.034 ms 2.90 ms
We also tested the time consumption of the TFHE and cuFHE decryption algo-
rithms. The decryption time of both schemes is ms level, and the difference can be
ignored. The time consumption of uFHE’s key generation algorithm is much higher
than that of TFHE. The main reason is that the GPU is only suitable for tasks with less
branches, large amounts of data, and delay-insensitive tasks. However, the key gen-
eration algorithm involves more logical branches, so the computational efficiency is
affected. On the other hand, GPU access memory has a large latency and requires high
data alignment. A single Warp accesses a group of memory, which can only read
consecutive and aligned 128 bits at a time. But if the memory accessed is discrete, the
efficiency is greatly reduced. Because a single Warp core cannot perform subsequent
operations when the data is not processed. A large number of memory read and write
operations are involved in the key generation algorithm, and the process of generating
the public key and private key depends on the synchronization of the results, which
resul in a slower speed.
In Table 1, we can see that the Nand gate operation time on the GPU platform is
approximately 4.5 times that on the CPU platform. Stream management and memory
management are important factors affecting the efficiency of the algorithm. In cuFHE,
additional streams are created, allowing kernel and memcpy operations to run in par-
allel, so the execution order is more flexible and improves execution efficiency. At the
same time, use the following methods to optimize data transfer efficiency: reduce the
amount and frequency of data transmitted between the host and the device; use page-
locked memory for higher bandwidth; reasonably arrange memory space, such as
frequently used data put into the cache, reduce data cross-thread exchange.
6 Conclusion
In this paper, we debugged the homomorphic encryption software library TFHE and its
corresponding GPU version cuFHE through experiments, verified their reliability and
analyzed their performance. The experimental results show that the speed of cuFHE is
about 4.5 times that of TFHE. The GPU can greatly improve the homomorphic running
speed of FHE algorithm.
References
1. Tang, D.-H., Zhu, S.-X., Wang, L., et al.: Fully homomorphic encryption scheme from
RLWE. J. Commun. 35(1), 173–182 (2014)
2. Huang, H., Gong, T., Chen, P., et al.: Secure two-party distance computation protocol based
on privacy homomorphism and scalar product in wireless sensor networks. Tsinghua Sci.
Technol. 21(4), 385–396 (2016)
3. Chen, Z.-G., Wang, J., Song, X.-X.: Survey on fully homomorphic encryption. Appl. Res.
Comput. 31(6), 1624–1630 (2014)
4. Gentry, C.: Fully homomorphic encryption using ideal lattices. In: STOC 2009, vol. 9,
pp. 169–178 (2009)
5. Zhou, T., Yang, X., Zhang, W., Wu, L., et al.: Efficient fully homomorphic encryption with
circularly secure key switching process. a a 1, 1
6. Gentry, C., Sahai, A., Waters, B.: Homomorphic encryption from learning with errors:
conceptually-simpler, asymptotically-faster, attribute-based. In: Advances in Cryptology–
CRYPTO 2013, pp. 75–92. Springer, Heidelberg (2013)
7. Chillotti, I., Gama, N., Georgieva, M., et al.: Faster fully homomorphic encryption:
bootstrapping in less than 0.1 s. In: Advances in Cryptology–ASIACRYPT 2016:
Proceedings of the 22nd International Conference on the Theory and Application of
Cryptology and Information Security, Part I, vol. 10031, Hanoi, Vietnam, 4–8 December
2016, pp. 3–33. Springer, Heidelberg (2016)
102 H. Yang et al.
8. Zhou, T.P., Yang, X.Y., Liu, L.F., et al.: Faster Bootstrapping with Multiple Addends.
Cryptology ePrint Archive, Report 2017/735 (2017). http://eprint.iacr.org/2017/735
9. Wang, W., Hu, Y., Chen, L., et al.: Accelerating fully homomorphic encryption using GPU.
In: High Performance Extreme Computing, pp. 1–5. IEEE (2013)
10. Dai, W., Doröz, Y., Sunar, B.: Accelerating NTRU based homomorphic encryption using
GPUs. In: IEEE High Performance Extreme Computing Conference, pp. 1–6. IEEE (2015)
11. Dai, W., Sunar, B.: cuHE: a homomorphic encryption accelerator library. In: International
Conference on Cryptography and Information Security in the Balkans, pp. 169–186.
Springer, Cham (2015)
12. Fast Fully Homomorphic Encryption Library over the Torus. https://github.com/tfhe/tfhe
13. Dai, W.: CUDA-accelerated Fully Homomorphic Encryption Library. https://github.com/
vernamlab/cuFHE
NADEEM: A Novel Reliable Data
Delivery Routing Protocol
for Underwater WSNs
Nadeem Javaid(B)

Abstract. In this research work, we propose three schemes: neigh-

bor node approaching distinct energy efficient mates (NADEEM), fall-
back approach NADEEM (FA-NADEEM) and transmission adjustment
NADEEM (TA-NADEEM). In NADEEM, immutable forwarder node
selection is avoided with the help of three distinct selection parame-
ters. Also, void hole is avoided using fallback recovery mechanism to
deliver data successfully at the destination. Moreover, transmission range
is dynamically adjusted to resume greedy forwarding among the network
nodes. The neighbor node is only eligible to become forwarder when it is
not a void node. Additionally, linear programming based feasible regions
are computed for an optimal energy dissipation and to improve network
throughput. Extensive simulations are conducted for three parameters:
energy, packet delivery ratio (PDR) and fraction of void nodes. Further,
an analysis is performed by varying transmission range and data rate for
energy consumption and fraction of void node. The results clearly depict
that our proposed schemes outperform the baseline scheme (GEDAR) in
terms of energy consumption and fraction of void nodes.
1 Introduction
In the last decade, UWSNs have gained a lot of attention from researchers
because of their potential to monitor underwater environment. UWSNs have
extensive variety of applications such as military defense, monitoring aquatic
environment, disaster prevention, oil/gas extraction, offshore exploration, com-
mercial and scientific purposes, etc. [1,2].
In underwater atmosphere protocols of terrestrial wireless sensor networks
(TWSNs) cannot work perfectly. The TWSN and UWSN have differences in
many aspects such as use of acoustic links instead of radio links, UWSNs topology
is more dynamic than TWSNs because nodes move freely with water currents
and change their position frequently, localization of nodes is difficult as compared
to TWSNs. In addition, UWSNs face many challenges like low bandwidth, high
propagation delay and high communication cost [3].
Various routing protocols are proposed to minimize the energy consumption
and avoid void occurrence [4,5]. To recover data from void node, fallback recovery
https://doi.org/10.1007/978-3-030-15035-8_10
104 N. Javaid
mechanism and transmission adjustment approaches are widely used. Hence, an

efficient protocol is required which can minimize energy consumption and also
be able to avoid void node occurrence.
Opportunistic routing is exploited to transmit data reliably by selecting mul-
tiple neighbor nodes at each hop. However, geographic routing paradigm causes
void hole problems because of the selection of only one node for data transmis-
sion. On the other hand, the opportunistic routing results in redundant trans-
mission and consumes more energy [6]. Thus, to achieve multiple objectives at
the same time, both schemes are combined to improve energy efficiency and
avoid void node occurrence in the network.
Motivated by above consideration, we propose neighbor node approaching
distinct energy efficient mates (NADEEM) and its two variant with special fea-
tures of void recovery using fallback and transmission adjustment approaches.
The NADEEM uses geographic and opportunistic routing for energy effi-
cient data communication. While in transmission adjustment NADEEM (TA-
NADEEM), the transmission of the void node is adjusted to bypass the void
region and forwards the data to the immediate available node. While fallback
recovery NADEEM (FA-NADEEM) uses backward transmission and finds a dif-
ferent routing path for successful data delivery at the destination.
2 Related Work and Problem Statement

Depth based routing (DBR) [7] only requires depth information to forward data
packets towards the sink with greedy approach. Each node forwards the data
packet based on depth parameter and also suppresses communication to avoid
redundant transmission via calculating holding time. The priority is assigned to
forwarder node solely on the basis of depth, however, it depends that packet is
new and not delivered at the destination. The DBR provides improved network
lifetime and high data delivery ratio. However, this greedy mechanism with one
parameter forces immutable selection of forwarder node due to which network
nodes get partitioned and resources remain under utilized.
An improved adaptive mobility of courier nodes in threshold-optimized DBR
(iAMCTD) [8] is a location free routing protocol, which is designed for time
critical applications. It provides an improved network lifetime, minimized end-
to-end delay with the help of mobile courier nodes. However, this scheme results
in low throughput, because of avoidance of the unnecessary transmission.
In balanced load distribution (BLOAD) [9], authors tackle the problem
of energy holes using the mechanism of fragmentation of data packet. The
fragments of every data are delivered via direct and multi hop transmissions
towards the destination. Moreover, nodes deployed away from the destination are
equipped with higher battery resulting in improved network lifetime. Authors
in weighting depth and forwarding area division-DBR (WDFAD-DBR) [10] con-
sidered the depth of current and next forwarding node in the direction of sink
deployed at the water surface. The information helped in avoiding void node
selection with less packet drop ratio. However, knowledge up to hops never guar-
antee reliable forwarder selection especially when the deployment is sparse.
NADEEM: A Novel Reliable Data Delivery Routing Protocol 105
An adaptive clustering habit ((ACH)2 ) is presented in [11], which proposes

the mechanism of free association of nodes with cluster heads (CHs). The CHs
are first elected on the premise of threshold value and then, an optimal num-
ber of CHs is selected based on the distance between each head node. In this
way data load managed among the head nodes which significantly minimizes
the energy consumption. Moreover, it reduces propagation distance and evades
back transmissions to reduce energy dissipation however, high communication
overhead in associating normal nodes with head nodes.
A hop-by-hop dynamic addressing based routing protocol for pipeline moni-
toring (H2-DARP-PM) is proposed in [12] for efficient forwarder node selection.
It uses dynamic addressing to obtain reliable and most effective forwarder node
in terms of energy consumption. Also, assigns dynamic hop address to each node
which contributes in delivering data at the destination. Although, the PDR is
enhanced, however, at cost of high energy consumption.
Cluster based sleep wake scheduling is performed by assuming initiator nodes
in [13]. The initiator nodes initiate communication and select the CHs. The CH
with high energy is selected to active mode while other nodes sent to sleep mode.
The transmission is resumed once head node is nominated and this decreases
energy consumption along with high lifetime and throughput. However, the
immutable selection of CH degrades the network performance. While authors
in [14] proposed an energy efficient cluster head selection protocol named parti-
cle swarm optimization (PSO-ECHS). To improve the performance of this algo-
rithm, various metrics including distance between the clusters, distance between
sinks and residual energy of nodes are taken into account.
To enable communication among the sensor nodes, an acoustic architecture is

required to illustrate the working mechanism of the network. Our proposed
schemes are based on multi sink architecture where nodes are randomly deployed
in the ocean and sink nodes at the surface of water [2] as shown in Fig. 1. It con-
sists of two types of nodes: relay nodes and anchored nodes [10]. The nodes
that are anchored are static at the bottom of the ocean whereas relay nodes
are deployed randomly in the acoustic environment. The relay and anchor nodes
use acoustic signal to carry information towards the destination. The anchored
nodes obtain the sensed information and forward it towards the sink located
at the surface with the help of relay nodes. Whereas, at the surface, sink nodes
relay data through radio signals for transmitting information at the base station.
It is assumed that packets received at a single sink are considered to be received
at all sink nodes successfully.
In our work, we made some assumptions because our focus was on the energy
efficiency and void avoidance. (i) A consideration has been made that every node
(relay, anchored and sink nodes) can obtain its coordinates with the help of local-
ization services [6]. (ii) The communication is symmetric, where communication
between any two random nodes results in the consumption of the same amount
106 N. Javaid
of energy. For instance, node i transmits data to node j with energy E and when
j sends data to i, the battery dissipation remains the same. It only changes when
the distance between two nodes vary. (iii) All acoustic nodes has the power to
adjust communication range autonomously and sink node can receive multiple
packets at the same time without any data loss or collision. (iv) The effect
of water currents is in the direction of horizontal, while in vertical direction
the movement is almost negligible. In the following paragraph an overview of
Thorp’s propagation model [2] has been presented to model the energy con-
sumption according to the distance between the source and the destination.
3.1 UWSNs Propagation Model
The path loss due to unhindered propagation route for a signal having frequency
f over a distance l is given as [15]:
A(l, f ) = ls α(f )l , (1)
where s is the spreading factor and α(f ) is the absorption coefficient. The geom-
etry of propagation is described using the spreading factor k and its values
for spherical, cylindrical and practical spreading are: s = 2, s = 1, and s = 1.5
respectively. The absorption coefficient α(f ), in dB/km for f in KHz, is described
by the Thorp’s formula [2] as:
10logA(l, f ) = s.10logl + l.10logα(f ). (2)
The common signal-to-noise ratio (SNR) over l is given as:
Energyb /A(l, f ) Energyb

SN R(l) = = , (3)
N oise0 N oise0 ls α(f )l
where Energyb and N oise0 are constants that reveal per bit energy transmis-
sion and noise power density on a non declining additive white gaussian noise
(AWGN) channel [3]. As in [16], rayleigh fading is used to model small scale
system where SNR has the following distribution probability:
∞
1 Y
ρl (Y ) = e− SN R(l) . (4)
0 SN R(l)
The probability of error is given as:

∞
ρe (l) = ρe (Y )ρl (Y )lY. (5)
0
Here, ρe (Y ) denotes the random modulation at a specific value of SNR and Y

is the probability of error. In this paper, we use the binary phase shift keying
(BPSK) modulation in which each bit is carried by a symbol [17]. The error
probability over l is taken from [18,20].

1 SN R(l)
ρe (l) = (1 − ). (6)
2 1 + SN R(l)
The data packet delivery probability for n bits is given by:
ρ(l, n) = (1 − ρe (l))n . (7)
4 Proposed Work
In this section, we describe our proposed schemes in detail. Initially, we will dis-
cuss the beacon message dissemination in the network for configuration. Addi-
tionally, neighbor node selection mechanism will be presented for successful data
communication. Moreover, an efficient mate selection is discussed to show its
effectiveness for reliable data transmission.
4.1 Dissemination of Beacon Message
The beacon dissemination is vital to enable the configuration of the nodes for
successful data communication. Once all nodes and sinks are deployed in the
network volume, collaboration is important to relay data in multihop fashion
from a source to destination. Thus, a beacon is transmitted by a sink node from
the surface of water which is also equipped with acoustic modem. This message
includes unique information like identification coordinates (x, y, z). The sinks
are equipped with positioning system and have the area facts of each other [2].
Similar is the case for sensor nodes. The beacon transmitted by sensor nodes con-
sists of sequence number, ID and x, y, z coordinates. The positioning system is
ineffective in underwater due to excessive frequency signal absorption [2]. There-
fore acoustic localization services are used for acquiring nodes information [6].
108 N. Javaid
Moreover to reduce the energy consumption and minimize the communication

overhead, the size of the beacon is reduced by only keeping information of node
ID and coordinates. After receiving the information, a node updates its entry
in the neighbor table. Beacons are generated to inform about the location of
the sensor node. With this knowledge, neighbor nodes can also access an energy
optimal forwarder node within the communication range.
4.1.1 Distinct Mates Selection

An optimal utilization of resources is always desired. In acoustic environment,
node battery is one of the most crucial asset that needs to be efficiently con-
sumed during the process of data communication. The void occurrence causes
network partition, thus an approach which can avoid the occurrence or bypass it,
is required. Therefore, we opt to select various forwarding mates from a neigh-
borhood [19,21]. The multiple neighbor selection concept is to ensure backup
when node with highest priority fails to deliver the desired data packet at the
respective destination. The following mathematical expression is used to find out
mates with desired features.
E(i)
M (i) = . (8)
(Trange × N ) + 1
In aforesaid equation, M (i) shows the mate selected for data communication
in the network. While i depicts the number of mates selected (i varies from 0
to N , where N could be any real positive integer). The right hand side of the
expression shows computation mechanism of priority factor. Here E(i) denotes
the energy of node i and it is multiplied with N number of neighbors to increase
its suitability because higher the neighbor number is, more will be backup avail-
able incase of transmission failure and 1 is added to avoid when N number is
0. The fundamental concept of this approach is to boost the data packet in the
direction of sinks on the surface. The selection of multiple neighbor nodes results
in less number of retransmissions of the data packets [18,20].
Moreover, one more factor is considered before nominating the forwarder
node, packet advancement (ADV) [18] to choose the forwarder node which leads
the packet towards the destination. The ADV is defined as the difference between
distance of a source node Si and destination node Di with difference of distance
between neighbor Y and Ni [2]. Where nj be the node which has a packet
to transmit and its neighbors set is represented as Nj (t) and set of sinks is
represented as Sj (t), where t is representing the time. Therefore, the neighbor
node set in geographic and opportunistic routing is given as:
Ni = {nk ∈ Nj (t) : ∃sv ∈ Sj (t)|Nd (nj , s∗i ) − Nd (nk , sv ) > 0}, (9)
where Si∗ is from set of sinks Si (t) and it is closest sink of node nj as:
s∗i = argmin∀sj ∈Sj (t) {Nd (nj , sj )}. (10)

4.1.2 Energy Efficient Mate Selection

The sender node set selection is based on normalized ADV (NADV) as proposed
in [2,3,20]. Using NADV, the most appropriate next hop forwarder mate from
the neighbor node set Ni is selected. The NADV relates the foremost economical
trade-off between the proximity and link cost to choose the priorities of the eligi-
ble nodes. So greater the NADV of node is greater will be its priority for selection
as next forwarder. For every next-hop forwarder node nf ∈ Ni , NADV is:
N ADV (nf ) = ADV (nf ) × ρ(lfj , n), (11)
here, ADV (nf ) = D(nj , s∗j ) − D(nf , s∗f ) is the nf ADV towards the closest sink,
s∗f ; lfj is the distance between source node nj and forwarder node nf and ρ(lfj , n)
is the probability of n bits over distance lfj and it is given in Eq. 7.
After the node set selection and next hop forwarder set selection assumes
fi ⊆ Ni is the set that is formed according to the priorities of NADV [2,20].
As the goal of geographic and opportunistic routing is to find the fi ⊆ Ni to
maximize the expected packet advance (EPA). The EPA of fi is formulated
in Eq. 11.
k
EP A(fi ) = Σd=1 N ADV (nd )Πj=0d−1
(1 − ρ(lij , n)). (12)
Ultimately, the set with maximum EPA is chosen as next forwarder node set.
The node with highest precedence is chosen from the forwarder set as a next hop
forwarder. It declares its location records to its neighbor node. If a node is not
decided as a next hop forwarder node for a longer period, its packet is discarded.
4.2 The FA-NADEEM

When the sparsity of the network increases, the void node occurrence probabil-
ity also rises. However, the lack of recovery mechanism make it vulnerable to
high packet drop ratio. Therefore, FA-NADEEM is used to ensure that packet
from void region must be recovered and delivered at the destination via an
alternate path. Figure 2 illustrates the mechanism of FA-NADEEM. It is evi-
dent when source node S transmits the data packet, it obtains the information
up to two hops, along with neighbor information to ensure the suitability of the
elected node to proceed with the data communication. There are four neighbor
nodes: n1, n2 n3 and n4. The first hop node is n2 because it is more effective to
relay packet in terms of energy and also helps in reducing the hop number. The
NADEEM further explores the neighbors of n2 and then n3 and so on. When
the n3 is reached, it encounter void region. Thus, fallback approach is adopted
in NADEEM and called FA-NADEEM. Figure 2 shows n3 declares it void and
instead of dropping the packet, it looks in its list based on the priority to ensure
it has to traverse least number of hops in recovery procedure. Similarly, n3 falls
back and transmits the data to n4 in its communication range. Then from n4,
the process of data forwarding is continued in the direction of sink deployed
at the water surface. In communication period, then fall back recovery proce-
dure is adopted to discover different routes which leads towards the destination.
110 N. Javaid
As soon as a node is found, greedy forwarding is resumed to save node battery.

Let’s assume a scenario depicted in Fig. 2, where node S forwards data packet
to the node n2 and when at node n3 it detects the void hole region then it uses
the fall back mechanism to forward the data packet towards node n4.
Fig. 2. Working mechanism of FA-NADEEM
4.3 The TA-NADEEM
The TA-NADEEM has the same steps of NADEEM except the adjustment of
transmission to avoid void hole problem which is illustrated in Fig. 3. During
communication TA-NADEEM computes the number of neighbors of the for-
warder node after adjusting the transmission range. Let us assume a scenario as
depicted in Fig. 3, where node S has to send a data packet towards the destina-
tion, it has no neighbor in its transmission range. Instead of opting for fallback
procedure, it simply adjusts the transmission range and delivers the data packet
directly to the immediate node comes into its adjusted communication range.
Fig. 3. Adjustment of transmission range in TA-NADEEM
When S fails to find a neighbor node, it adjusts its transmission range as we

mentioned earlier in system model that every node has the ability to adjust the
transmission range incase of void occurrence to avoid packet drop. The imme-
diate forwarder is n1, S acquires the information about the neighbors of n1
to ensure that selected node is not a void. Thus, it looks into the neighbor
table of n1 and finds that it has two neighbors at higher depth than n1. Thus,
n1 is selected as the effective forwarder node to continue with the process of
data routing. If there is no neighbor in the communication vicinity of n1, then
further adjustment in transmission range is done until a suitable forwarder is
located. Additionally, it updates the information in packet queue if there is no
report regarding the packet delivery. It helps in avoiding redundant data packets
transmission resulting in less energy consumption which prolongs the network
lifespan. Moreover, in future, nodes look into the neighbor table and avoids the
void occurrence. In this way extra energy consumed in transmission adjustment
is accommodated.
To evaluate the performance of proposed schemes (NADEEM, FA-NADEEM

and TA-NADEEM), we conducted extensive simulations against existing scheme
GEDAR [3]. The detail is given in the upcoming subsections.
5.1 Performance Control Parameters
To perform simulations, the calibration of input parameters which directly affect

the performance of the network system. The simulations are run for 10 times and
average is plotted using the line graphs. Moreover, number of nodes is randomly
deployed in a 3 dimensional (3D) network volume of 1500 m × 1500 m × 1500 m
with an increment of 50 nodes to 150 until it reaches to 450 nodes. Additionally,
the transmission range is set to 250 m for comparison and it varies in transmission
analysis from 150 m to 350 m. Also, data rate is set to 16 kbps for performance
evaluation against existing scheme GEDAR [3], whereas to analyse the deviation
in energy and void occurrence is observed on 32 kbps, 64 kbps and 128 kbps. In
addition, payload of the data packet is set to 150 bytes. The values of energy
consumption are: Pt = 2 W, Pr = 0.1 W and Pi = 10 mW for transmission,
reception and idle energies, respectively [3].
5.2 Fraction of Local Maximum Nodes
Figure 4 illustrates the ratio of void nodes of NADEEM and its variants against
the GEDAR. It is evident from the results that void ratio decreases with the
increase in node number. The highest transmission failure ratio can be observed
of FA-NADEEM which falls significantly up to 250 nodes. Then again, after the
node density of 300, the amount of void node occurrence decreases to a certain
level, which remains consistent till 450 nodes. The decrease in ratio is because
of the fall back mechanism of FA-NADEEM which looks for an alternative route
112 N. Javaid
from all nodes in the communication range to recover and deliver data packet
successfully at the destination.
NADEEM has less void appearances at the end of the node increment because
it does not have transmission adjustment, fall back and depth adjustment mech-
anisms (Fig. 4). This help in saving node battery ultimately leading to less nodes
and more number of alive nodes. Thus, when the sparsity of the nodes decreases,
the probability of void occurrence also decreases.
GEDAR and TA-NADEEM show almost identical trend in rise and fall of
void proportion during the nodes communication. Although, the values are dif-
ferent and at the end TA-NADEEM beats the baseline scheme because of direct
transmission adjustment, instead of exchanging beacon message to find out the
different route (Fig. 4). When different route is not available then depth of the
node is adjusted, these aspects consume a lot of energy and suddenly node bat-
tery depletes. Whereas, TA-NADEEM adaptively adjusts its transmission power
and bypass the void node successfully.
From 150–200 nodes GEDAR shows sudden decrease in fraction of void nodes
(Fig. 4). As GEDAR considers movement of void node towards its neighbor node
without checking whether the neighbor nodes have enough forwarder nodes or
not, therefore it has high fraction of void nodes compared to proposed schemes.
0.5
NADEEM
0.45 GEDAR
FA-NADEEM
0.4
Fraction of local maximum nodes
TA-NADEEM
0.35
0.3
0.25
0.2
0.15
0.1
0.05
0
150 200 250 300 350 400 450
Number of nodes
Fig. 4. Comparison of fraction of void nodes
5.3 Packet Delivery Ratio
Figure 5 shows that the energy tax gradually decreases with the increase in
density of network. The highest PDR is achieved by FA-NADEEM due to its
ability of finding data route from an alternate node within the communication
vicinity. When a trap of void node occurs, instead of dropping the packet, the
source node proceed with fall back procedure and looks in 360◦ to resume the
data transmission process.
NADEEM depicts moderate success ratio of packets transmission compared
to other schemes in Fig. 5. The performance is neither high nor low, it operates
on static parameters, however, the forwarder selection is not based on single

parameter of depth.
On the other hand, the performance regarding success ratio of GEDAR and
TA-NADEEM is identical in Fig. 5. The energy consumption is high in both of
the methodologies because in GEDAR, depth adjustment in vertical direction
which consumes more energy as compared to simple recovery methods. Similarly,
the TA-NADEEM opts towards the adjustment of transmission range, the energy
consumption depends on the distance between the source and destination. When
the transmission power is adjusted, it requires more energy to transmit the
signal to overcome the issues of attenuation and fading in dynamic acoustic
environment.
0.35
NADEEM
GEDAR
0.3
FA-NADEEM
TA-NADEEM
0.25
Packet delivery ratio
0.2
0.15
0.1
0.05
0
150 200 250 300 350 400 450
Number of nodes
Fig. 5. Comparison of PDR
5.4 Energy Consumption

The comparison of energy consumption in delivering single packet at the desti-
nation of all the schemes is depicted in Fig. 6. The minimum energy consumption
can be observed from the graphical representation of TA-NADEEM and slightly
high of GEDAR. The energy consumed for one packet is almost 0.02 J of TA-
NADEEM, whereas, the node battery dissipation in GEDAR is high than all the
proposed schemes which very small number of nearly 0.12 J.
The energy consumption of FA-NADEEM is comparatively more than the
other two proposed schemes and almost the same with GEDAR (Fig. 6). The
reason of more energy is of alternate route which increases the route length and
resulting in more energy requirement. Although, it achieves better PDR (Fig. 5),
however, still the utilization of limited resource battery is considerably higher.
On the contrary, TA-NADEEM shows minimum energy dissipation in deliv-
ering a single data packet. It happens due to only energy required in adjusting
power and factors like message exchange, depth adjustment are not involved.
Thus significant amount of energy is saved in TA-NADEEM. Whereas NADEEM
has initially low energy requirements, but, when node density increases, the col-
lision rate of data packets also maximizes resulting in high packet loss.
114 N. Javaid
1
NADEEM
0.9 GEDAR
FA-NADEEM
Energy per data packet per node (J)

0.8 TA-NADEEM
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
150 200 250 300 350 400 450
Number of nodes
Fig. 6. Comparison of energy consumption
6 Conclusion
In this paper, we have proposed NADEEM, FA-NADEEM and TA-NADEEM
routing protocols to enhance the energy efficiency and minimize the fraction
of void node occurrence. The NADEEM minimizes the energy consumption by
avoiding immutable forwarder selection. It also achieves higher packet delivery
ratio. Whereas, FA-NADEEM avoids successfully the void node through send-
ing data from the alternate route. However, the use of alternate path increases
the number of hops which resulted in more energy consumption. Moreover, it
has higher throughput as compared to NADEEM. On the other hand, when
FA-NADEEM fails to find an another forwarding path then transmission of
NADEEM is adjusted to ensure data delivery at the destination node. Addition-
ally, TA-NADEEM looked for an immediate forwarder node and its neighbors to
ensure that selected node is not void. It has low network throughput, however,
minimum energy consumption for one packet. The results are validated through
extensive simulations. The outcome clearly shows that proposed schemes outper-
formed the baseline existing scheme (GEDAR) in terms of energy consumption
and void avoidance.
References
1. Javaid, N., Jafri, M.R., Ahmed, S., Jamil, M., Khan, Z.A., Qasim, U., Al-Saleh,
S.S.: Delay-sensitive routing schemes for underwater acoustic sensor networks. Int.
J. Distrib. Sens. Netw. 11(3), 532676 (2015)
2. Coutinho, R.W.L., Boukerche, A., Vieira, L.F.M., Loureiro, A.A.F.: Geographic
and opportunistic routing for underwater sensor networks. IEEE Trans. Comput.
65(2), 548–561 (2016)
3. Noh, Y., Lee, U., Lee, S., Wang, P., Vieira, L.F.M., Cui, J.-H., Gerla, M., Kim,
K.: Hydrocast: pressure routing for underwater sensor networks. IEEE Trans. Veh.
Technol. 65(1), 333–347 (2016)
4. Kheirabadi, M.T., Mohamad, M.M.: Greedy routing in underwater acoustic sensor
networks: a survey. Int. J. Distrib. Sens. Netw. 9(7), 701834 (2013)
5. Han, G., Jiang, J., Bao, N., Wan, L., Guizani, M.: Routing protocols for underwater
wireless sensor networks. IEEE Commun. Mag. 53(11), 72–78 (2015)
6. Coutinho, R.W.L., Boukerche, A., Vieira, L.F.M., Loureiro, A.A.F.: GEDAR: geo-
graphic and opportunistic routing protocol with depth adjustment for mobile
underwater sensor networks. In: 2014 IEEE International Conference on Com-
munications (ICC), pp. 251–256. IEEE (2014)
7. Yan, H., Shi, Z.J., Cui, J.: DBR: depth-based routing for underwater sensor net-
works. In: NETWORKING 2008 Ad Hoc and Sensor Networks, Wireless Networks,
Next Generation Internet, pp. 72–86 (2008)
8. Javaid, N., Jafri, M.R., Khan, Z.A., Qasim, U., Alghamdi, T.A., Ali, M.: iAMCTD:
improved adaptive mobility of courier nodes in threshold-optimized DBR protocol
for underwater wireless sensor networks. Int. J. Distrib. Sens. Netw. 10(11), 213012
(2014)
9. Azam, I., Javaid, N., Ahmad, A., Abdul, W., Almogren, A., Alamri, A.: Balanced
load distribution with energy hole avoidance in underwater WSNs. IEEE Access
5, 15206–15221 (2017)
10. Yu, H., Yao, N., Wang, T., Li, G., Gao, Z., Tan, G.: WDFAD-DBR: weighting
depth and forwarding area division DBR routing protocol for UASNs. Ad Hoc
Netw. 37, 256–282 (2016)
11. Ahmad, A., Javaid, N., Khan, Z.A., Qasim, U., Alghamdi, T.A.: (ACH)2 : routing
scheme to maximize lifetime and throughput of wireless sensor networks. IEEE
Sens. J. 14(10), 3516–3532 (2014)
12. Abbas, M.Z., Bakar, K.A., Ayaz, M., Mohamed, M.H., Tariq, M.: Hop-by-hop
dynamic addressing based routing protocol for monitoring of long range underwater
pipeline. KSII Trans. Internet Inf. Syst. (TIIS) 11(2), 731–763 (2017)
13. Yu, S., Liu, S., Jiang, P.: A high-efficiency uneven cluster deployment algorithm
based on network layered for event coverage in UWSNs. Sensors 16(12), 2103
(2016)
14. Javaid, N., Rasheed, M.B., Imran, M., Guizani, M., Khan, Z.A., Alghamdi, T.A.,
Ilahi, M.: An energy-efficient distributed clustering algorithm for heterogeneous
WSNs. EURASIP J. Wirel. Commun. Network. 2015(1), 151 (2015)
15. Heidemann, J., Stojanovic, M., Zorzi, M.: Underwater sensor networks: applica-
tions, advances and challenges. Phil. Trans. R. Soc. A 370(1958), 158–175 (2012)
16. Noh, Y., Lee, U., Han, S., Wang, P., Torres, D., Kim, J., Gerla, M.: DOTS: a prop-
agation delay-aware opportunistic MAC protocol for mobile underwater networks.
IEEE Trans. Mob. Comput. 13(4), 766–782 (2014)
17. Stojanovic, M., Preisig, J.: Underwater acoustic communication channels: propa-
gation models and statistical characterization. IEEE Commun. Mag. 47(1), 84–89
(2009)
18. Yang, H., Liu, B., Ren, F., Wen, H., Lin, C.: Optimization of energy efficient trans-
mission in underwater sensor networks. In: Global Telecommunications Conference,
GLOBECOM 2009, pp. 1–6. IEEE (2009)
19. Climent, S., Sanchez, A., Capella, J.V., Meratnia, N., Serrano, J.J.: Underwater
acoustic wireless sensor networks: advances and future trends in physical, MAC
and routing layers. Sensors 14, 795–833 (2014)
20. Noh, Y., Lee, U., Wang, P., Choi, B.S.C., Gerla, M.: VAPR: void-aware pressure
routing for underwater sensor networks. IEEE Trans. Mob. Comput. 12(5), 895–
908 (2013)
21. Yadav, S., Kumar, V.: Optimal clustering in underwater wireless sensor net-
works: acoustic, EM and FSO communication compliant technique. IEEE Access
5, 12761–12776 (2017)
An Architecture for Distributed
Ledger-Based M2M Auditing for Electric
Autonomous Vehicles
Dragos Strugar(B) , Rasheed Hussain, Manuel Mazzara, Victor Rivera,

Ilya Afanasyev, and JooYoung Lee
Innopolis University, Universitetskaya str. 1, 420500 Innopolis, Russia

{d.strugar,r.hussain,m.mazzara,v.rivera,
i.afanasyev,j.lee}@innopolis.ru
Abstract. Electric Autonomous Vehicles (EAVs) promise to be an effec-

tive way to solve transportation issues such as accidents, emissions
and congestion, and aim at establishing the foundation of Machine-to-
Machine (M2M) economy. For this to be possible, the market should
be able to offer appropriate charging services without involving humans.
The state-of-the-art mechanisms of charging and billing do not meet this
requirement, and often impose service fees for value transactions that
may also endanger users and their location privacy. This paper aims at
filling this gap and envisions a new charging architecture and a billing
framework for EAV which would enable M2M transactions via the use
of Distributed Ledger Technology (DLT).
1 Introduction
Vehicles, once considered mechanical boxes and used as a primary source of com-
muting, are now transforming into smart cars. With the massive introduction of
software, modern cars are equipped with embedded sensors, actuators and com-
putational capabilities, including communication and storage. This surge in the
computation, communication, and automotive technologies gave rise to Intelli-
gent Transportation System (ITS) realized through Vehicular Ad hoc NEtworks
(VANET) [13]. Simultaneously, Electric Propulsion Vehicle (EPVs or EVs) were
introduced to address the drastic increase of global greenhouse gas (GHG) emis-
sions. Despite the numerous regulations that were made to address the ecological
problem, technological solutions are also necessary to be used in synergy. If we
also consider recent advances in the field of Artificial Intelligence (AI) we have all
the instruments to potentially reduce human error in transportation systems by
building self-driving vehicles. They are capable of autonomously steering, nav-
igating, vehicle agile maneuvering, making decisions, foreseeing potential acci-
dents and acting better than human drivers in some critical situations [14,31].
Having such machines at our disposal could result in numerous benefits, includ-
ing, but not limited to, fewer accidents, reduced traffic congestions, and enable-
ment for so-called Mobility-as-a-Service (MaaS) [16]. Let’s consider these topics
in details.
https://doi.org/10.1007/978-3-030-15035-8_11
An Architecture for Distributed Ledger-Based M2M Auditing for EAVs 117
Accidents. In the United States, motor vehicle crashes are one of the leading
causes of death, as approximately two million collisions happen every year [1,2].
A system which would eliminate the human error in the vehicle collisions would
result in a smaller number of vehicle crashes.
Traffic Congestion. It is the cause of numerous negative effects including
delays, air pollution and increased chance of collisions. This problem has not
been solved yet and is one of the main concerns for big cities and megalopolises.
Research has shown that only 5% of Autonomous Vehicles on our roads can
significantly reduce the stop-and-go traffic waves that can arise due to lane
changing [27].
MaaS. Mobility-as-a-Service is a concept that delivers users’ transportation
needs through a service provider (SP) [16]. The core idea is to have a fleet of
Self-driving vehicles (SDVs) which can perform transportation services for their
users. This could lead to the decline in car ownership and private car usage and
the expansion of the Sharing economy. Companies like Volkswagen are very inter-
ested in this type of transportation systems, as it has the potential to change the
automobile industry [17]. With the increasing popularity of EV, the need of hav-
ing a sufficient charging infrastructure appears. Currently, the most widespread
option is the use of Charging Stations (CSs) to deliver electrical energy from the
power source to the EV, ensuring the appropriate and safe flow of electricity.
CSs are the main interface between the vehicle, the utility and the user [6]. This
means that the service providers have enabled users to charge their vehicle by
plugging it into the CS, and then paying for the services provided through the
use of credit/debit cards. Although this may seem convenient for human use,
they are not suitable for the Machine-to-Machine (M2M) economy [3]. The vehi-
cles of the future will be driverless, requiring the absence of human interaction
in the charging process (this is a typical scenario for MaaS). In addition, such an
approach leads to fees imposed often by trusted third-party authorities. Finally,
they violate the bidirectional auditability and are not privacy-aware [15]. A new
system which would solve these issues is desirable.
DLTs and the Tangle. A new decentralized architecture and distributed com-
puting paradigm called the Blockchain has gained significant popularity. It rep-
resents an immutable ledger system implemented in a distributed way which
eliminates the need for a trusted third-party [19]. By using technologies relying
on distributed ledgers we could increase the transparency of the whole system
while maintaining the high level of its security. The research of DLT-related use
cases has been actively conducted for multi-agent systems [7,18], and one par-
ticular area which could be disrupted by the use of this innovative technology
is Internet-of-Things (IoT). The main reasons why IoT could benefit from the
DLT are M2M value transactions, increased security and the automation of pro-
cesses. However, technical limitations of the current blockchain architecture do
not allow it to be broadly used in this industry, primarily due to high transac-
tion fees and scalability issues [24]. Machines will need to make micropayments
between themselves, but in order to make them feasible, transaction fees should
be very minimal, preferably non-existent. In addition, transaction confirmation
118 D. Strugar et al.
times should be very short for the network to be scalable and ready to accept new
transactions. Recently, the Tangle has been introduced [21]. It is the next step
in the evolution of Distributed Ledger Technologies and is an underlying tech-
nology of the IOTA cryptocurrency. Based on a concept of a Directed Acyclic
Graph (DAG), Tangle still serves as a distributed database, but it has three
main advantages: it is scalable, involves no fees and allows for offline transac-
tions. These characteristics of the Tangle and its cryptocurrency make it possible
to become the backbone of the emerging Machine Economy.
With these concepts in mind, the cryptocurrency which uses the Tangle,
IOTA, aims to become the backbone of the emerging Machine Economy, by
allowing machines to trade resources (electricity in our case) and services with
each other without the involvement of the intermediary.
Outline of the Paper. This paper is focusing on the problem of charging Elec-
tric Autonomous Vehicles without human mediation using Distributed Ledger
Technologies. The problem is open and its solution may have significant impact
in shaping the society of the future. In this Section we have presented the general
scenario of EAV and Distributed Ledger Technologies, with particular attention
to IOTA, the cryptocurrency we are proposing for the payment method. In
Sect. 2 we present the overall Architecture with discussion of the technological
stack while in Sect. 3 we show the paradigmatic use case and what features the
platform should exhibit. Section 4 introduces the technological complexities of
the project, Sect. 5 discusses how to protect EAVs and CSs from malicious cheat-
ing in serving and billing, and Sect. 6 brings out some reflections and plans for
the future. In this paper we define the vision and the architecture of our frame-
work. The formal design of automotive system that can be applied to this case,
has been described in [10].
2 Platform Architecture
In this section we describe the overall system architecture and its organization
as well as the proposed technologies. Figure 2 depicts the architecture consisting
of three major layers:
1. Physical
2. Network
3. Services
2.1 Physical Layer
This layer encapsulates all the hardware components embedded in the CS used
for sensing and gathering information about the charging process, as well as
communicating with external entities. The core parts of this layer are:
• kWh meter
• Main Controller (MC)
• EVSE Controller (EVSEC)
Fig. 1. Vision of unmanned EAV charging
The main sensor in the CS is the kWh meter as it measures the amount
of electricity consumed by the EAV. Along with the current electricity market
price, this data would be constantly used to calculate the total sum that the
vehicle would pay to the service provider. The benefit of such an approach is
that users get the exact amount of electricity they have paid for.
Moreover, knowing the current electricity rate could help us in adjusting the
maximum power the EAV gets charged with, especially in the case of multiple
CS using the same power source. The device used to perform such an operation is
the Smart Electric Vehicle Supply Equipment (EVSE) [26], previously referred
to as EVSEC. This could deliver electricity in a controlled (smart) way, and
contribute to the development of the Smart Grid [4,25].
Lastly, we use MC to transfer the information gathered by components of
the Physical layer to higher levels of the architecture through the use of a com-
munication protocol specifically designed for IoT, Message Queuing Telemetry
Transport (MQTT).
2.2 Network Layer

Since our scenario is composed of autonomous, dynamic decision-making CSs
and vehicles, we suggest the communication between the entities in the system
to be entirely peer-to-peer (P2P). Therefore, peers (EAV and CS) should have
equal privileges [30]. In order to allow for P2P communication, we consider DLTs
as a possible solution. In particular, we choose IOTA’s Tangle as decentralized
database that stores both data and value transactions. Communication mod-
ule built in such a way would benefit from the wide range of possibilities for
networking methods, as parties could transfer data and value either offline or
Fig. 2. System architecture
online, depending on the protocol used. Since our architecture involves constant
M2M communication with subscription behavior, we consider MQTT connectiv-
ity protocol to be the most appropriate choice for inter-device communication
[12]. In our case, MQTT would ensure rapid transfer of both data and value with
high throughput.
In order to allow value transactions in such a system, a more secure solution
is needed. Recently, a bi-directional payment channel has been introduced which
meets the requirements stated above and conforms to the MQTT protocol by
nature. It is called the Flash Channel (FC) [9], and it allows both resource and
service providers to get paid for their services while they are performing them.
FCs enable real-time streaming of transactions, and allow machines to branch-
off the main network, transmit data and value via whichever protocol they find
suitable (MQTT in our case), and then merge the results with the main IOTA
Tangle. This means that there is no need to wait for transactions to be approved
by the IOTA network. By using it, only the starting and ending transactions of
the Flash channel will be published to the main network.
Once machines have finished transacting in FC, the final transaction, denoted
as tf , becomes ready to be published on the Tangle. When this happens, tf has to
verify two not yet approved transactions. It does so by performing computational
work which is used to verify if the transaction is trusted or not. This process is
called the Proof-of-Work (PoW). Similarly, other transactions tk =f , tl=k =f will
find and verify tf . In order for any transaction tx =f to be approved in the Tangle
network, two incoming transactions ty and tz have to verify it using their PoW.
Such an approach removes the need for miners, and therefore enables feeless trans-
actions. In addition, the network benefits from new transactions as they will verify
two unapproved ones and thus improve the scalability of the system [21].
Through the use of the Tangle, FCs and MQTT protocol for message transfer,
the Network Layer would be able to transfer sensor data from the Physical Layer
to the Services Layer rapidly and securely.
2.3 Services Layer

On the top of the architecture sits the Services Layer which makes use of the
two layers described above to deliver services to both consumers and service
providers:
• Charging services for EAVs
• Data Insights for Service Providers
As end-consumers of charging services are not humans, but rather EAVs,
we envision the application used for payment of the services provided, to be
embedded inside the vehicle itself. This application would make the use of Global
Positioning System (GPS) and Artificial Intelligence (AI) to locate the most
appropriate Charging Stations. In addition, it would be responsible for initiating
the charging process, broadcasting the amount of electricity needed and paying
for services provided using its IOTA wallet. Lastly, once the charging process is
completed, the application would close the payment channel and record the new
wallet balance on the main Tangle network. This way, EAV would have its own
wallet, and truly embrace the M2M Economy by automating the whole process.
On the other hand, service providers would benefit from the fact that they
can analyze the data collected in real-time and gain useful insights. As state-of-
the-art NoSQL databases enable low latency, the data collected from the Physical
and Network Layers in the process of off-tangle Flash transactions could be used
for Machine Learning (ML) scenarios such as graphical demos, predictions and
live billing analysis.
3 Unmanned EAV Charging

Our vision demands investment on the creation of Electric Roads (ER), meaning
roads with electric strips that would allow charging on-the-go [11]. Wireless
power transfer (WPT) has been around since Tesla [28]. In such a scenario,
EAV would be able to charge without stopping at the station and without the
need of any human interaction. The payment framework that we envision is
based on IOTA. The ideal scenario is pictorially represented in Fig. 1. While the
wireless method provides numerous benefits, and we consider it to be a more
suitable long-term solution for EAV charging, its mass-scale deployment is not
yet feasible at the moment [29]. Alternatively, regular Charging Stations (CSs)
should be considered.
When the vehicles’ Battery Management System (BMS)1 indicates that the
vehicle needs charging, the application embedded in the EAV starts searching for
1
BMS is a combination of sensors, controller, computation and communication hard-
ware with software algorithms designed to estimate the battery percentage of the
EV [5].
the most suitable CS nearby. It does so by performing AI algorithms based on

both Cloud data gathered from various charging points and the current vehicle’s
location determined by the GPS.
Once arrived at the Charging Point, EAV authenticates with the CS by
fetching its wallet address and establishing a secure payment channel, i.e. FC.
The address fetched by the vehicle will be used for value transfer once FC has
done all the preparatory steps. When the authentication step is completed and
the FC is established, both parties have to deposit equal amount of IOTA into a
multi-signature address controller by both parties. This step is mandatory as it
would prevent parties from refusing to continue to sign the transactions. When
these deposits are confirmed, FC does not interact with the network until the
channel is closed. After both devices have successfully deposited funds inside a
FC, EAV broadcasts the amount of electricity it needs to the CS. Networking
layer processes the request and sends the data to the MC of the Physical Layer.
When the CS gets this information, the charging process starts.
While EAV is getting the electricity from the CS, MC is using data from
kWh meter to record the amount of electricity consumed. This data is then
sent to the EAV’s application via MQTT protocol, and the consumer pays the
calculated sum through the FC. The transactions are stored in a local (preferably
NoSQL) database with low latency. This enables service providers to perform
ML algorithms on this data and gain useful insights in real-time. In addition,
service providers could also benefit from live-billing analysis and charging process
monitoring.
When EAV gets the amount of electricity it has asked for, or when the
charging process is manually stopped, FC takes the latest state and spreads
the remaining tokens across transacting parties. After each party signs the final
bundle, it is then attached to the main network, i.e. the Tangle.
EAV’s application records the end of the charging process, and sets its state
to inactive.
4 Research Challenges
The capabilities of IOTA are unprecedented and it is poised to revolutionize the
current micro-economy; however, in the domain of connected and autonomous
cars industry, there are challenges that need to be addressed. In this section, we
focus on the general IOTA-related and EAV-specific challenges.
Adaptation. IOTA is still in its infancy and at the moment, it is struggling with
growing the Tangle. Nevertheless, IOTA undoubtedly provides many benefits as
opposed to blockchain, for instance no transaction fees; however, the adaption
of IOTA among businesses is still a challenge and needs to address different
issues such as best practices of cryptography (the absence of which may lead
to security loopholes), conflict resolutions, approval of new members, and so
on. It is, at least at the moment, premature to guarantee its adoption among
the consumers and the businesses. In other words, it is still speculative what
the future will bring for IOTA, but the signs are convincing that IOTA will
become popular for micro-transactions provided that the underlying challenges

are addressed.
Performance. Mining in blockchain is a time-consuming procedure with the
growth in the network which renders it unsuitable for micro-transactions. IOTA -
the successor of blockchain, on the other hand, performs better when the network
grows. However, at the moment IOTA is not as adopted as much as blockchain
and it will need time to grow. Therefore, we will have to wait and see how
IOTA evolves in the years to come. Meanwhile, IOTA needs other means of
increasing the performance as compared to its other competitors in the market.
Furthermore, the performance challenge is also indirectly related to the consumer
trust to the IOTA which means that unless consumers are satisfied, the network
will not grow and subsequently, the performance will not increase.
Security. Currently IOTA has a “Coordinator” that is managed by IOTA Foun-
dation and it protects users from malicious activities such as DoS and other
attacks. However, when the IOTA network grows, IOTA Foundation is planning
to remove the coordinator and completely rely on the Tangle itself. The cur-
rent attacks that target IOTA are mainly based on the DAG where malicious
nodes with enough computational power can create a sub-tangle which will cause
the IOTA network to approve this malicious transaction. To deter this kind of
attack, a strong reputation and validation mechanism is essential. Furthermore,
in the current state, the coordinator keeps malicious nodes away from trying
to launch different kinds of attacks such as the one discussed above and others
such as parasitic chain attacks (where malicious transactions keep trying unless
they are approved by the network even though with a very low probability) and
other splitting attacks. However, when the network grows and the coordinator
is removed from the IOTA network, it is imperative to address these security
issues beforehand.
Trust Management. Reputation is based on the number of transactions
through a particular node in Tangle. The greed for higher score will let mali-
cious nodes to try as many transactions as they can. As aforementioned, this
phenomenon could lead to the approval of malicious transactions. Therefore
efficient and effective authorization and authentication mechanisms aided with
reputation management are essential for IOTA.
Trust in Coordinator. The holistic centralized trust in IOTA Foundation
where they guarantee the security and validity of transaction could also endan-
ger the efficacy of IOTA. This phenomenon is also related to technical issues
with the underlying “nuts and bolts” of the IOTA system, i.e. cryptographic
protocols. It is always recommended to use the existing best practices crypto-
graphic protocols; however, IOTA uses its own home-grown hash function curl.
It is worth noting that there has been found vulnerabilities in the hash function
used by IOTA which may endanger the trust in IOTA network. Furthermore,
another problem with IOTA coordinator is that it is a closed source which gives
no room for the application developers to maneuver. In order for IOTA network
to excel, it has to consider these challenges before they decide to remove the
coordinator from the network.
5 Threat Model
This section discusses how the proposed system would protect EAVs and CSs
from malicious cheating during the settlement of payments. One key characteris-
tic that makes the system we proposed in this article different is that it effectively
solves the Guest-Host problem (GHP) This problem presents the case of a guest
who wants to use a hotel room for some number of days and does not want to
get billed for the service used it. On the other hand, the host should not get paid
less than determined and agreed by both parties. In our case, we consider EAV
as the guest, and CS as the host of the GHP problem.
We consider three main scenarios: (A) Guest tries to get more charge than
it has paid for, (B) Host tries to bill the user more than it is agreed, and (C)
Either side tries to benefit from the interruption of the charging process.
A. Guest tries to get more charge than he/she has paid for
Problem. The most common scenario occurs when the guest gets the service as
promised, but does not want to pay the settled amount of money. For this very
reason, state-of-the-art electric CSs mostly require users to have special service
cards tied to their bank accounts and often passports/national IDs2 . This way
the CS can ensure that the user will get billed. Thus, scenarios where EAVs first
get the charge and then pay, are not convenient for service providers and are
seldom used.
Solution. Our approach does not assume that someone should pay first. Instead,
we look at the process of charging and billing as token streaming. Key idea is
to have a system where both charging and billing are happening at the same
time. That is, while EAV is getting the charge, it is paying for these services.
Our solution leverages IOTA, and their bidirectional Flash Channels (FCs) for
this specific use case. As we have discussed, FCs enable instantaneous, high-
throughput transactions. This fact is crucial in our scenario. Potentially, to be
entirely sure that the system is fair, the number of transactions per second
can be set. It would ensure that the users get exact amount of charge for the
amount they pay. Additionally, IOTA, by its nature, is privacy-aware, as IOTA
transactions do not include any personal information3 .
Thus, the guest (EAV) can not get more charge than it has paid for. And
additionally, the whole process of billing has numerous benefits, including, but
not limited to: trust, fairness for both parties, transparency and privacy.
2
But it is a prime example of Privacy violation - why service providers are able to
see users’ detailed private information?
3
IOTA transaction contains: transaction hash, signature message fragment, address
of the recipient, value, timestamp, bundle, nonce, and a few other properties that
reveal nothing about the identity of transacting parties.te.
B. Host tries to bill the user more than it is agreed
Problem. While the state-of-the-art approach does in fact solve one aspect of
the GHP (host does get paid), it can be easily seen that this approach is not
perfect for the guest. First, EAV exposes its private information, then pays for
the service, and in the end, has to trust the CS that it will deliver the promised
charging services. And not only that, CS can hide the current market price, and
bill users way more than they should be billed. Therefore, users must be able to
see the price they are getting the electricity for.
Solution. Our proposed scheme solves this problem very effectively. At each
point in time, users are billed with the amount of money which corresponds to
the amount of electricity they get, as there are no fees. The whole process of
charging and billing is transparent for the guest of the GHP. If the host tries
to bill the guest more than it is agreed, that malicious activity can be easily
identified, and the host could have reporting contract violation warning.
C . Either side tries to benefit from the interruption of the charging

process
Problem. One notable case occurs when either the guest or the host decides to
quit the charging process by force. In traditional charging station systems, there
are very few ways to deal with such a situation. What service providers do, is
employ people to regulate the whole process that degrades the M2M concept.
Solution. As in the previous two cases, payment channels handle also provide a
very elegant solution. Either side could terminate the charging process by force,
and the only consequence would be that the channel closes. When that happens,
CS constructs a bundle that takes the balance of the channel, and divides it
amongst the channel’s users accordingly. These bundles are then attached to the
network, and neither side benefits from this interruption.
6 Conclusions
The emergence of Intelligent Technologies is pervading every moment of our
life. While cars themselves may externally look similar to decades ago (apart
from design trend consideration), a dramatic revolution is happening inside the
cabin. This revolution could be described by the single word: software. Soft-
ware is becoming dominant, at times pervasive, of everyday existence. This cer-
tainly demands for high attention on privacy and ethical issues, however expands
our possibilities beyond the unthinkable. The same way that we have intelli-
gent portable devices, we have intelligent buildings [22,23] that can significantly
improve our lives or the lives of people with impaired capacities [20], and so we
are aiming at having intelligent cars. In this paper we discussed the synergy of
three fundamental technologies: Autonomous Vehicles, Electric Cars and cryp-
tocurrencies. What we propose is their cooperation in order to implement M2M
charging with realization of a billing framework for EAVs via IOTA technol-
ogy. Future work will involve the full realization of the vision here described, at
first with a computer graphical simulation, and finally with a full scale develop-
ment. After these three phases have been completed, we are going to present the
project to authorities in order to upgrade the project to a larger scale. Further
refinements could include a microservice-based architecture, aimed at increasing
scalability [8], which is already part of our research agenda. At a larger scale, the
vision could be expanded foreseeing a marketplace of individuals and companies
selling out energy in excess or energy generated precisely for this purpose. Once
the technology will be in place, this would create a specific market run on the
idea of cryptocurrency as a value-exchanger in a scenario where cars can drive
and stop for charge in a completely autonomous manner.
References
1. Critical reasons for crashes investigated in the national motor vehicle crash causa-
tion survey. Technical report, U.S. Department of Transportation, 1200 New Jersey
Ave, SE, Washington, DC, 20590, USA, February 2015. https://crashstats.nhtsa.
dot.gov/Api/Public/ViewPublication/812115
2. Traffic safety facts, 2015 - a compilation of motor vehicle crash data from the
fatality analysis reporting system and the general estimates system. Technical
report, U.S. Department of Transportation, 1200 New Jersey Ave, SE, Wash-
ington, DC, 20590, USA (2015). https://crashstats.nhtsa.dot.gov/Api/Public/
ViewPublication/812384
3. Bojkovic, Z., Bakmaz, B., Bakmaz, M.: Machine to machine communication archi-
tecture as an enabling paradigm of embedded internet evolution. In: International
Conference on ACE, pp. 40–45 (2014)
4. van den Brink, H.: How ElaadNL built a PoC charge station running fully on IOTA,
and IOTA only (2017). https://medium.com/@harmvandenbrink/how-elaadnl-
built-a-poc-charge-station-running-fully-on-iota-and-iota-only-e16ed4c4d4d5
5. Cheng, K.W.E., Divakar, B.P., Wu, H., Ding, K., Ho, H.F.: Battery-management
system (BMS) and SOC development for electrical vehicles. IEEE Trans. Veh.
Technol. 60(1), 76–88 (2011)
6. Cities, C.: Plug-in electric vehicle handbook for public charging station hosts. US
Department of Energy Publication No. DOE/GO-102012-3275 (2012)
7. Danilov, K., Rezin, R., Kolotov, A., Afanasyev, I.: Towards blockchain-based robo-
nomics: autonomous agents behavior validation. In: 9th International Conference
on Intelligent Systems. IEEE (2018)
8. Dragoni, N., Lanese, I., Larsen, S.T., Mazzara, M., Mustafin, R., Safina, L.:
Microservices: how to make your application scale. In: 1st Convergent Cognitive
Information Technologies, pp. 95–104 (2017)
9. Freiberg, L.: Instant & feeless flash channels (2017). https://blog.iota.org/instant-
feeless-flash-channels-88572d9a4385
10. Gmehlich, R., Grau, K., Iliasov, A., Jackson, M., Loesch, F., Mazzara, M.: Towards
a formalism-based toolkit for automotive applications. In: 1st FME Workshop on
Formal Methods in Software Engineering (FormaliSE) (2013)
11. Guardian: How world’s first electrified road charges moving vehicles (2018). www.
youtube.com/watch?time continue=56&v=VZNHZnyxCm8
12. Hunkeler, U., Truong, H.L., Stanford-Clark, A.: MQTT-S—a publish/subscribe

protocol for wireless sensor networks. In: 3rd International Conference on Commu-
nication Systems Software and Middleware and Workshops (COMSWARE), pp.
791–798. IEEE (2008)
13. Hussain, R., Son, J., Eun, H., Kim, S., Oh, H.: Rethinking vehicular communica-
tions: merging VANET with cloud computing. In: 4th International Conference on
CLOUDCOM, pp. 606–609. IEEE (2012)
14. Hussain, R., Zeadally, S.: Autonomous cars: research results, issues and future
challenges. IEEE Commun. Surv. Tutor., 1 (2018)
15. Hussain, R., Kim, D., Nogueira, M., Son, J., Tokuta, A.O., Oh, H.: PBF: A
new privacy-aware billing framework for online electric vehicles with bidirectional
auditability. CoRR abs/1504.05276 (2015). http://arxiv.org/abs/1504.05276
16. Jittrapirom, P., Caiati, V., Feneri, A., Ebrahimigharehbaghi, S., Alonso-González,
M., Narayan, J.: Mobility as a service. Planning 2(2), 13–25 (2017)
17. Jungwirth, J.: Digitalization and mobility for all (2017). https://youtu.be/
x2iBLkfyQuA
18. Kapitonov, A., Lonshakov, S., Krupenkin, A., Berman, I.: Blockchain-based proto-
col of autonomous business activity for multi-agent systems consisting of UAVs. In:
Research, Education and Development of Unmanned Aerial Systems (RED-UAS),
pp. 84–89. IEEE (2017)
19. Nakamoto, S.: Bitcoin: a peer-to-peer electronic cash system (2008)
20. Nalin, M., Baroni, I., Mazzara, M.: A holistic infrastructure to support elderlies’
independent living. In: Encyclopedia of E-Health and Telemedicine. IGI Global
(2016)
21. Popov, S.: The tangle. cit. on p. 131 (2016)
22. Salikhov, D., Khanda, K., Gusmanov, K., Mazzara, M., Mavridis, N.: Jolie good
buildings: Internet of Things for smart building infrastructure supporting concur-
rent apps utilizing distributed microservices. In: 1st Convergent Cognitive Infor-
mation Technologies, pp. 48–53 (2016)
23. Salikhov, D., Khanda, K., Gusmanov, K., Mazzara, M., Mavridis, N.: Microservice-
based IoT for smart buildings. In: 31st Advanced Information Networking and
Applications Workshops (WAINA) (2017)
24. Schiener, D.: Introduction to IOTA cryptocurrency (2017). https://blog.iota.org/
a-primer-on-iota-with-presentation-e0a6eb2cc621
25. Siano, P.: Demand response and smart grids—a survey. Renew. Sustain. Energy
Rev. 30, 461–478 (2014). http://www.sciencedirect.com/science/article/pii/
S1364032113007211
26. Stegen, M.: SmartEVSE - smart electric vehicle charging controller (2014). https://
github.com/SmartEVSE/smartevse
27. Stern, R.E., Cui, S., Monache, M.L.D., Bhadani, R., Bunting, M., Churchill, M.,
Hamilton, N., Haulcy, R., Pohlmann, H., Wu, F., Piccoli, B., Seibold, B., Sprinkle,
J., Work, D.B.: Dissipation of stop-and-go waves via control of autonomous vehi-
cles: Field experiments. CoRR abs/1705.01693 (2017). http://arxiv.org/abs/1705.
01693
28. Tesla, N.: Experiments with Alternate Currents of Very High Frequency and their
Application to Methods of Artificial Illumination. pp. 336–337. The Electrical Engi-
neer, NY, 18 March 1891
29. Williams, J.: Transport innovation of the week: electric charging lanes
(2017). https://makewealthhistory.org/2017/01/23/transport-innovation-of-the-
week-electric-charging-lanes/
30. Yuan, Y., Wang, F.Y.: Towards blockchain-based intelligent transportation sys-
tems. In: IEEE 19th International Conference on Intelligent Transportation Sys-
tems (ITSC), pp. 2663–2668. IEEE (2016)
31. Zubov, I., Afanasyev, I., Shimchik, I., Mustafin, R., Gabdullin, A.: Autonomous
drifting control in 3D car racing simulator. In: 9th IEEE International Conference
on Intelligent Systems. IEEE (2018)
Is QUIC Quicker Than TCP?
A Performance Evaluation
Saurabh1 , Satya Prakash1 , Pranav Kumar Singh1,2 , Sunit Kumar Nandi1,3 ,

and Sukumar Nandi1(B)
1
Department of CSE, Indian Institute of Technology Guwahati,
Guwahati 781039, India
sukumar@iitg.ac.in
2
Department of CSE, Central Institute of Technology Kokrajhar,
Kokrajhar 783370, Assam, India
3
Department of CSE, National Institute of Technology, Arunachal Pradesh,
Papum Pare 791112, India
Abstract. TCP is a well-known protocol used by many web applica-

tions. However, to address the shortcomings of TCP including connec-
tion setup time and congestion response time, Google developed QUIC
(Quick UDP Internet Connection). QUIC, a UDP based protocol which
is similar to TCP + HTTP/2 + TLS has several advantages over mod-
ern TCP including highly reduced connection initiation time, improved
congestion control and better connection migration. This paper explores
how performance of TCP and QUIC is affected under different network
and control parameters like bandwidth, added delay, packet loss, spikes
and number of flows.
Keywords: QUIC · TCP · UDP · TLS · HTTP2.0
1 Introduction
The internet since it’s invention has become the most efficient and effective ways
to communicate. Nowadays we depend on the internet for a myriad of things,
including business, online transactions, education, inventory, marketing, banking
and communication etc. It is important that we ensure a fair, smooth and efficient
internet all over the globe. However, since online traffic is constantly changing
at any given time or place, it is difficult to ensure the same.
There is a need for globally accepted internet protocols that will ensure effi-
cient and fair use of internet between any two machines. TCP [9] is one such
protocol that has been around 1982. TCP is a widely used protocol that ensures
fair and efficient network with built in effective congestion control mechanisms
for traditional web applications. However, with increase in number of users, web
traffic, content size and interactive web applications, TCP has failed to maintain
optimal network conditions.
QUIC [8] is a protocol developed by Google to address many of the issues
encountered with TCP. As shown in Fig. 1, it is an application layer protocol
https://doi.org/10.1007/978-3-030-15035-8_12
130 Saurabh et al.
that runs on the top of UDP [10] protocol. With QUIC, Google has tried to
provide some key features including reduced connection establishment time in
0-RTT(average), better connection migration, improved congestion control and
reduction of HOL blocking. Google is still rethinking some design decisions like
error recovery features as previously implemented FEC didn’t prove to be too
useful. QUIC implements TCP like properties in the application layer rather
than in the transport layer.
Fig. 1. QUIC - where it fits
Google chrome uses QUIC by default and all the Google servers employ this
protocol. Referring [2] since its introduction in 2013, traffic using this protocol
increased significantly, currently approximately 58% mobile traffic and 64% video
traffic uses QUIC (Fig. 2). Many of the Google owned web applications like
youtube, gmail, drive etc. have already started using QUIC for their internet
traffic. However, it is still in experimental stage and there is a lot of room for
improvement. A dummy QUIC server and QUIC client, only able to be used for
integration testing and therefore not scalable has been released publicly online
for anyone to use and toy with [6].
Fig. 2. QUIC video traffic share for mobile networks

Is QUIC Quicker Than TCP? 131
1.1 Contribution
We explore and compare performance analysis of QUIC and TCP. We designed

tests that would analyze performance of TCP and QUIC under added delay,
packet loss, spikes, number of flows and variable bandwidth. The results were
compared in terms of total transfer time, throughput, average bandwidth used,
overhead in bytes and fairness index. These tests were run under emulated net-
work conditions on a wired network.
1.2 Paper Organization
Our paper explores QUIC and TCP’s performance over various control param-
eters, and analyzing the raw data to compare them. Section 2 discusses related
work. Section 3 gives a detailed discussion of TCP and QUIC. Section 4 talks
more about the experiment setup and methodology used to achieve results in this
paper. Section 5 discuss the performance analysis and results obtained. Section 6
presents conclusions based on these results and finally Sect. 7 discuss future work.
2 Related Work
There has been no significant work done over performance of QUIC over con-
gested networks under variable control parameters. Megyesi et al. [12] focuses
on page load times over QUIC as compared to TCP + HTTP1.1 under lab con-
ditions, Carlucci et al. [3] also analyses QUIC in the terms of page load times.
Hussein et al. [1] observed QUIC performance over 3D web. Quentin et al. [5]
compares performance of MPTCP with their own implementation of multipath
QUIC under lossy links. Lychev et al. [11] prioritizes cryptographic schemes
present in QUIC. Sarah et al. [4] focuses on page load times under QUIC with
added packet loss rates.
On paper QUIC looks very powerful and successful in improving on TCP’s
limitations, but it is still an immature technology carrying only limited func-
tionality unlike TCP which can be used for file sharing, remote shell access and
network routing. QUIC might still be too far to being a replacement for TCP, but
given enough time we may see many of its successful features being implemented
in TCP.
3 Background
This section briefs built-in mechanisms provided with both TCP and QUIC.
3.1 Transport Control Protocol (TCP)
This section uses TCP Reno as reference to briefly describe various built-in
concepts of the TCP protocol.
132 Saurabh et al.
Connection Establishment. TCP uses 3-way handshake for connection estab-

lishment. The client first sends a SYN packet to the server, requesting connection
initiation. The server then responds by acknowledging this request by returning
a SYN-ACK packet. The client finally sends an ACK packet acknowledging the
SYN-ACK packet and thus establishing a connection between the two machines.
Sliding Window. As shown in Fig. 3 the Sliding window algorithm imple-

mented by TCP ensures that the receiver doesn’t get overwhelmed by more data
than it can process. The sender keeps on sending data till either the receivers
window gets fully filled or it receives acknowledgement from the receiver at which
point the window slides forwards allowing more data to be sent in.
Fig. 3. TCP’s sliding window
Slow Start. TCP ensures that no new connections steal bandwidth away from
existing connections. This is done by exponentially increasing the size of sliding
window.
Congestion Avoidance. When Transmit Window size (TWS) grows beyond

Slow start threshold (Ssthresh), TCP starts limiting the growth of window
size. This is done whenever TCP encounters packet losses and interprets it as
congestion.
Fast Recovery. This improvement was added to TCP Reno and called New
Reno when it was discovered that most fast re-transmit actions occur during
mild congestion. Instead of using slow start TCP increases congestion window
size for every duplicate ACK received after the third one until we encounter a
non-duplicate ACK. The congestion window size and slow start threshold are
then reduced to reflect packet loss.
3.2 Quick UDP Internet Connections (QUIC)
Connection Initiation Time. QUIC handshakes on average require 0-RTT as

opposed to 1–3 RTT required for TCP + TLS. The first time a client connects
to a server, it must perform 1-RTT and in response server replies with a source
address token and servers certificates. Next time this client wants to connect to
the same server, it can use cached credentials to establish a connection in 0-RTT
(Fig. 4).
Fig. 4. Zero RTT connection establishment time
Multiplexing. QUIC is designed from the ground up for multiplexing opera-

tions. A QUIC packet is composed of multiple frames containing streams of data
independent to each other and identified by unique Stream ID.
A TCP frame is seen as continuous stream of data, which means if a single
packet is lost, no streams on that connection can make progress even if that
packet is sitting in buffer. QUIC can handle these scenarios with ease as streams
can continue receiving packets independent of other streams.
Forward Error Correction (FEC). QUIC can recover from lost packets
without having to wait for retransmission of this packet using FEC packets.
QUIC compliments a group of packets with a FEC packet which contains parity
of packets in the FEC group. The sender may decide beforehand whether to use
FEC or not to optimize link conditions.
Connection Migration. QUIC connections are represented by a 64 bit con-

nection ID that is randomly generated by the client. This means that a QUIC
connection can be reused even when a client changes it’s IP address. QUIC
performs cryptographic validation of clients and therefore allowing them the
continued use of a connection. In, contrast TCP uses a 4 bit tuple containing
IP address to represent a connection, and therefore cannot reuse a connection
when IP changes.
134 Saurabh et al.
Congestion Control. Currently QUIC uses a reimplementation of TCP Cubic

along with providing richer information to the congestion control algorithm.
Every packet whether it is original or retransmitted, carries a new unique
sequence number which allows a QUIC sender to distinguish ACKs for retrans-
missions from ACKs for originals. QUIC can allow for precise RRT calculations
as all ACKs also carry the delay experienced between sending and receiving of
the ACK packets.
QUIC’s ACK frames support up to 256 NACK ranges, implying QUIC is
more resilient as compared to TCP when it comes to reordering.
4 Methodology
For TCP, this paper uses an apache2 server for server side whereas wget is used
to download files over HTTP/2+TLS over the client side. For QUIC, quic server
and client are used which uses iproute2 and netem to send the file from server
to client. The architecture is shown in Fig. 5.
Fig. 5. Experiment architecture
A 10.4 MB pseudo-random file is generated which is used for testing pur-

poses. The protocols used to get this file are TCP and QUIC. We simulate the
environment in localhost and use ‘tc netem’ and ‘tbf’ to configure local loop-
back interface. The mtu is set to a max size of 1500 to better simulate real life
conditions and IPv6 is disabled so as to ensure smooth working of QUIC.
For calculating fairness index we use Jain’s Fairness index [7] with number of
flows varying over 1, 2, 4 and 8 flows. Among these one flow is set to use either
TCP or QUIC protocol while every other flow uses TCP protocol.
Both QUIC and TCP servers were configured with valid encryption certifi-
cates and the experimental platform used was as follows
– Intel(R) Core(TM) i5-6500 CPU @ 3.20GHz.

– Quad core processor.
– 12GB of RAM.
– Ubuntu 16.04 LTS (64 bit).
– Linux 4.15.0-20-generic.
– GCC 7.3.
– Python 2.7.12.
The various control parameters were varied over different values and for each
set of parameter values, the tests were run for a total of 5 times. The bandwidth
rates were 100 Mbps 40 Mbps and 5 Mbps while delay rates were 10 ms, 50 ms
and 100 ms, packet loss rates were 0%, 0.5%, 1.0% and 5%, finally either spikes
were present or they were absent. The results are then plotted using python class
‘matplotlib’.
5 Analysis
The project ran in 3 phases.
5.1 Generate Raw Data
Generate the parameters from argument parsing. Configuration of local loopback

interface for every parameter. Data is captured with ‘tcpdump’ for TCP protocol
and ‘quic client’ for QUIC protocol. This data is then saved in a ‘./dirty’ folder.
5.2 Data Analysis
Clean the dirty data and extract only needed information like timestamp and
bytes in ‘./clean’ folder. Average the data results for same instances of test, on
average each test is run for 5 times by default.
5.3 Visualization
We generate plots for performance bandwidth, overhead and time vs delay, band-
width, packet loss.
The following observations were made from obtained results From Fig. 6
we observe, QUIC generally performed better than TCP when packet losses
were small, but for higher packet losses there was no significant difference in
throughput. We can interpret from Fig. 7 that, under high bandwidth and delay
QUIC will perform much better than TCP. From Fig. 12 it can be inferred that,
QUIC performs generally better than TCP but has to make a trade on overhead
Fig. 6. Throughput vs packet loss under 100 ms delay and 100 Mbps bandwidth
136 Saurabh et al.
Fig. 7. Throughput vs delay under 5% packet loss and 100 Mbps bandwidth
Fig. 8. Time series graph under 5 Mbps bandwidth 0% packet loss and 100 ms delay
Fig. 9. Time series graph under 100 Mbps Fig. 10. Throughput comparison of TCP
bandwidth 1% packet loss and 100 ms vs QUIC when jitter is present 1
delay
size. From Figs. 8 and 9 it can be seen that, QUIC is more stable TCP under
favourable conditions. From Figs. 10 and 11 we can infer that TCP may outper-
form QUIC if jitter is present. Finally from Figs. 13 and 14 we observe, QUIC is
very unfair when competing for bandwidth as opposed to TCP.
Fig. 11. Throughput comparison of Fig. 12. Overhead comparison with

TCP and QUIC when jitter is present 2 50 ms delay and 0.5% packet loss
Fig. 13. TCP fairness vs no. of flows Fig. 14. QUIC fairness vs no. of flows
6 Conclusions
QUIC performed similar to TCP when packet losses were involved, this is because
FEC has been currently disabled and needs to be explicitly specified in the header
to be used and therefore QUIC interprets packet losses the same way TCP does
and re transmits them.
It is evident that QUIC transfers happen faster than TCP transfers which can
be credited to high throughput and efficient bandwidth usage. QUIC is generally
more stable than TCP, this may be because of high NACK size of 256 for QUIC,
which can continue in-taking packets even when previous packets were dropped.
Even though QUIC uses same congestion control algorithm as TCP Cubic,
still QUIC almost always uses more than it’s fair share of bandwidth when
competing with TCP flows. This may be because of large congestion window
size of QUIC and it’s tendency to increase its window size more aggressively
than TCP, which allows it to use bandwidth available more greedily.
We observed that TCP can outperform QUIC when jitters is present, this
shows that QUIC is still incomplete and not perfect. We did not observe any
significant differences in QUIC and TCP even though theoretically QUIC has
way better stats. All in all these results should be taken with a grain of salt,
since they were performed at home over restricted network and not by Google
on their servers where they have millions of users to test on.
138 Saurabh et al.
QUIC is still lacking in many ways which is evident from Google continually
trying to improve and better it, particularly error recovery has a huge scope for
improvement since it is practically disabled right now.
7 Future Work
For future work it might be interesting to explore the following aspects. Ana-
lyzing stream multiplexing in QUIC and observing its effects on load times of
HTTP pages. Evaluation and comparison of TCP and QUIC during connection
migration. Performance evaluation of QUIC over a wireless network. Comparing
performance of QUIC and TCP with other protocols like MPTCP and SCTP.
Proposing schemes for further optimization of QUIC protocol.
References
1. Bakri, H., Allison, C., Miller, A., Oliver, I.: HTTP/2 and QUIC for virtual worlds
and the 3D web? Procedia Comput. Sci. 56, 242–251 (2015)
2. Bouzas, S.: Why the meteoric rise of google quic is worrying mobile opera-
tors. https://2nuzlzatrf-flywheel.netdna-ssl.com/uploads/Chart-for-25Jan18-blog.
jpg. Accessed 01 Dec 2018
3. Carlucci, G., De Cicco, L., Mascolo, S.: HTTP over UDP: an experimental investi-
gation of QUIC. In: Proceedings of the 30th Annual ACM Symposium on Applied
Computing, pp. 609–614. ACM (2015)
4. Cook, S., Mathieu, B., Truong, P., Hamchaoui, I.: QUIC: better for what and for
whom? In: IEEE International Conference on Communications (ICC 2017) (2017)
5. De Coninck, Q., Bonaventure, O.: Multipath QUIC: design and evaluation. In:
Proceedings of the 13th International Conference on emerging Networking Exper-
iments and Technologies, pp. 160–166. ACM (2017)
6. Google: Playing with QUIC. https://www.chromium.org/quic/playing-with-quic.
Accessed 01 Dec 2018
7. Jain, R.: Fairness: how to measure quantitatively? ATMF/94-0881 (1994)
8. Langley, A., Riddoch, A., Wilk, A., Vicente, A., Krasic, C., Zhang, D., Yang, F.,
Kouranov, F., Swett, I., Iyengar, J., et al.: The quic transport protocol: design and
internet-scale deployment. In: Proceedings of the Conference of the ACM Special
Interest Group on Data Communication, pp. 183–196. ACM (2017)
9. Lin, F., Li, X., Li, W.: Research on TCP protocol in wireless network and net-
work simulation. In: 4th International Conference on Wireless Communications,
Networking and Mobile Computing, WiCOM 2008, pp. 1–4. IEEE (2008)
10. Liu, X., Cheng, L., Bhargava, B., Zhao, Z.: Experimental study of TCP and UDP
protocols for future distributed databases (1995)
11. Lychev, R., Jero, S., Boldyreva, A., Nita-Rotaru, C.: How secure and quick is
quic? Provable security and performance analyses. In: 2015 IEEE Symposium on
Security and Privacy (SP), pp. 214–231. IEEE (2015)
12. Megyesi, P., Krämer, Z., Molnár, S.: How quick is QUIC? In: IEEE International
Conference on Communications (ICC), pp. 1–6. IEEE (2016)
PRAN: A Provenance Based Model
and Prototype to Strengthen Authentication
Rajidi Satish Chandra Reddy(&) and Srinivas Reddy Gopu
Tata Consultancy Services, Hyderabad, India

{rajidis.reddy,srinivas.greddy}@tcs.com
Abstract. In this paper, we present the provenance based authentication

(PRAN) model, and prototype that aims to strengthen password-based authen-
tication of a system—ensuring legitimacy of a user’s identity by detecting
anomalous user logins without the need for other obtrusive authentication fac-
tors. The PRAN model utilizes authentication provenance along with traditional
methods. We evaluate the prototype with password stealing/sharing use cases.
We observe that the PRAN prototype detects anomalous logins, and also
legitimate logins thereby preventing applications from allowing anomalous user
logins’.
1 Introduction
Provenance is defined as information about entities, activities, and people involved in

producing a piece of data or thing, which can be used to form assessments about its
quality, reliability, or trustworthiness [1]. Historically, it has been mostly used in
domains such as scientific, healthcare, workflow, chemistry, biology. In such domains,
there is a need to perform data analysis and elevate trust worthiness of data. Thus,
provenance access control (PAC)—securing provenance data—emerged to elevate
trustworthiness of the underlying data [2–4]. Recently, there has been considerable
amount of research work done on provenance based access control (PBAC) that
focuses on the usage of provenance data to secure the underlying data [5–9]. Prove-
nance data can provide utilities such as pedigree information, usage tracking, and
object versioning. Using the provenance data utilities, a security mechanism can have
flexible control capabilities that can enhance security of a system.
1.1 Motivation and Contributions

In spite of the well-known problems [10], password-based authentication is still widely
used because of security, deployability, and usability [11]. The recently emerged bio-
metric authentication has been the alternative for conventional “userid/password”
authentication because it is more secure, unobtrusive, and convenient. However, while
physical bio-metric is intrusive and has some problems during capture phase, behav-
ioral bio-metrics is still evolving, and has accuracy issues [15] as it stands today. The
password-based authentication is going to be used in future, unless, there is widespread
adoption of its alternatives. Hence, we attempt to strengthen the password-based

https://doi.org/10.1007/978-3-030-15035-8_13
140 R. S. C. Reddy and S. R. Gopu
authentication by leveraging provenance utilities to attack two problems: minimize

password stealing/sharing; minimize the need for intrusive authentication factors such
as token, OTP, and bio-metric.
Our paradigm’s novelty lies on the premise that differentiating anomalous login
attempts as a result of password sharing/stealing without compromising on the
usability, convenience, and intrusive aspects. We achieve this by leveraging authen-
tication provenance—utilizing causal relationships formed over the entities involved in
the users’ login events.
Our contributions are:
• A user study on authentication methods. This is discussed in the Problem statement
section.
• A provenance aware security model based authentication provenance to ensure
legitimacy of a user without the need for multi-factor authentication. This is dis-
cussed in the Provenance based Authentication model section.
• A prototype based on the model that serves as a proof of concept for detecting
password sharing/stealing scenarios. This is discussed in the prototype section.
• A policy specification mechanism that allows user specifies authentication policy in
terms of authentication provenance abstractions. This is discussed in the prototype
section.
• A Policy evaluation mechanism that detects authentication anomaly scenarios by
evaluating policies applicable to an authentication context. This is discussed in the
prototype section.
The rest of paper is organized as follows. We describe the problem in Sect. 2,
related work in Sect. 3, then we present the provenance based authentication model in
Sect. 4, and prototype in Sect. 5, its evaluation in Sect. 6, and limitations and chal-
lenges in Sect. 7, and finally we provide conclusions and future directions.
2 Problem Statement
In this hyper-connected world, enterprises still prefer to use password-based authenti-

cation mechanisms. It is coupled with single sign on feature to make more convenient
for users. However, attacker, who manages to steal password, will succeed to login into
all associated applications configured with single sign on. In password-based authen-
tication, the system assumes that the user who submitted the credentials is the credential
owner, and verification is done only using the authentication factor—password. Here,
the problem is system’s inability to determine actual credential owner (Identification).
This problem cannot be solved, even if the password’s strength is increased.
Apart from password stealing, we want to highlight another problem “password
sharing”. Users engage in password sharing using various means that are out of control
of a security mechanism. In an enterprise, usually, users engage in password sharing to
perform: elevated privilege tasks; delegation [23]. Even though, password sharing is
used for convenience, it leads to security implications—information leakage, identity
theft, privacy. In a customer environment, a user may perform unauthorized functions
resulting in bad reputation, fines, and loss of business, and litigation.
PRAN: A Provenance Based Model and Prototype to Strengthen Authentication 141
We conducted a user study using our internal collaborating platform to understand

the users’ pulse on the authentications methods. Our user study results indicate that:
MFA is more obtrusive, requires additional hardware, prone to time out issues, and
lacks usability; Password-based authentication is still a widely used, and will survive in
future, if it is improved in usability, security, and convenience aspects. These results
justify our reasoning to attempt strengthening password based authentication. So, we
intend to address password-based authentication problems based on a different
premise—strengthening authentication in an unobtrusive way as opposed to leveraging
multi-factors that are obtrusive, inconvenient, and hard to deploy. We have come up
with an idea to confirm the users’ identity utilizing authentication provenance.
3 Related Work
Our research work is focused more on provenance based security, and its application in
the area of authentication. Over the last few years, in the provenance related literature,
there have been some works done in the context of PAC: preserving integrity and
confidentiality [5, 6].
Recently, there has been considerable work done on PBAC. It has been recognized
that dependency relationships captured in a provenance aware system (PAS) [13] can
be used to control access to data. The work in [5] designed a basic PBAC model that
supports dynamic separation of duty, work flow. There have been discussions on access
control languages for PBAC [12], provenance aware access control policies [7], and
extension of PBAC model in [8]. To the best of our knowledge, there is no work in the
literature that addresses authentication problems using provenance utilities.
In the literature, there are some works that addresses password authentication
problems. Password stealing is addressed by strengthening passwords: using graphical
user interfaces [20] on the premise that humans remember images more than textual
passwords; proposed strong password protocols, and safe password entry interfaces
[21]; user authentication protocols [22]. These works are limited in usability aspects
due to obtrusiveness.
The work in [11] proposes a frame work for selecting an active authentication
factor based on contextual factors to balance security and usability of authentication. It
addresses the intrusive aspects of MFA but does not deal with the improving a way to
confirm user’s identity. The technique NIAS [14] uses knowledge of user to understand
the user’s identity using probabilistic methods based on users’ past access. Our work
deals with detecting potential anomalies as a result of password sharing/stealing based
on authentication provenance.
The behavioral bio-metrics [15] comes close to our work because of unobtrusive
way of capturing keystrokes, mouse dynamics to verify the claim of a user. However,
behavioral bio-metrics is new, and a lot of work needs to be done to overcome accuracy
issues, and has open challenges [17]. Our work is different from behavioral bio-metrics
in that we capture information about entities that are involved in a user authentication,
and use that information for evaluating the current authentication context based on
authentication policies that are specified in terms of abstractions. Existing literature in
password solutions has not paid much attention to password sharing problem.
4 Provenance Based Authentication Model
In this section, we present the PRAN model, and its components.

Let U be a set of users such that uɛU is uniquely represented; Let S is a set of
systems such that sɛS is uniquely represented; AP is a set of applications such that
apɛAP is uniquely represented; Let AC be a set of activities where acɛAC is uniquely
represented. User u may use a system s to login aclogin to a web application or a native
application which can be represented as
8u e U; s e S; ap e AP; ac e AC : u ! s ! ap1::i: ð1Þ
Let UA, SA, APA, ACA, CONA are a set of user, system, application, activity, and
contextual attributes such that uaɛUA, saɛSA, apaɛAPA, acaɛACA, and conaɛCONA
uniquely represented. Let AR is a set of authentication requests such that
8areAR: ar ! ðua, sa, apa, aca, cona): ð2Þ
Let AD is a set of authentication decisions such that AD ! {“allow”, “deny”,

“allow with additional factor”}. AVP is an authentication validation point which
intercepts login requests and forwards ar to authentication evaluation point
(AEP) component. The Fig. 1 demonstrates the PRAN model.
4.1 Authentication Provenance

Let PABS is a set of provenance abstractions, and ET is a set of extraction techniques,
PR is a set of parameters such that
8pabePAB, eteET, prePR:pab ! etðprÞ: ð3Þ
defines that each abstraction is mapped to an extraction technique with an input

parameter pr. The Fig. 2 is an authentication graph that demonstrates the relationships
of the login events of the users “Alice”, Bob”.
The provenance generation component captures authentication provenance for the
pre-defined actions. The captured provenance is stored in the authentication provenance
store. We describe the authentication provenance model below:
Let AG is a set of agents, EN is a set of entities, AC is a set of activities in
provenance model such that
8ueU, ageAG:ag ! u: ð4Þ
8seS, apeAP, eneEN:en ! ðs _ ap): ð5Þ
8ac e AC: ac ! ðacLogin _ acBrowse): ð6Þ

Let P be a set of predicates that define relation between AG, EN, AC. Let PATTR be
a set of attribute type, value pairs applied to AG, EN, AC as a triple such that
8pattrePattr,ageAG,eneEN,aceC:ag ! pattr:type ! pattr:Val,

ð7Þ
en ! pattr:type ! pattr:Val, ac ! pattr:type ! pattr:Val
Fig. 1. PRAN model Fig. 2. Authentication provenance graph
We adopt following PROV-DM [1] relationships: WasGeneratedBy(WGB),

WasAttributedTo(WAB), Used(U), WasDerivedFrom(WDF) such that WAB, WGB,
WDF, U ɛ P.
We use following statements ag ← WAT ← e, e ← WDF ← e, ac ← U ← e,
ac ← WGB ← e.
4.2 Policy Model

Let P be a set of authentication policies such that
8p e P : p ! ðrl e RL, APPLCR) ð8Þ
where RL—a set of rules specified in terms of provenance abstraction pabs—is

represented as
8rl e RL, ua e UA, sa e SA,apa e APA, aca e ACA, cona e CONA, pabsePABS: rl !
ðua, sa, apa, aca, cona, pabs):
ð9Þ
APPLCR is applicability criteria of the policy.

5 Prototype
We have developed the prototype to demonstrate the PRAN model. We have used
Apache jena1, prov tool2, and WSO2 Balana3 XACML java libraries:Apache jena for
capturing, querying the authentication provenance; prov tool java libraries for showing
the graphical representation of the authentication provenance; XACML [18] for policy
specification and evaluation. The Fig. 3 demonstrates the PRAN prototype architecture.
Fig. 3. Provenance based Authentication architecture
5.1 Authentication Validation Point (AVP)

The authentication validation point (AVP) will intercept all user logins and forwards
them to the policy engine. It will create an authentication context from the data
available that is understandable to the policy engine. The policy evaluation server will
evaluate the authentication context against applicable authentication policies.
Authentication policies are created by the security administrators specifying criteria
using provenance abstractions which will be a basis for detecting anomalies. The AVP
intercepts the user’s login and forms an AR with the user, system, activity, and con-
textual attributes.
1
https://jena.apache.org/.
2
http://lucmoreau.github.io/ProvToolbox/.
3
https://github.com/wso2/balana.
5.2 Authentication Server

The authentication server facilitates the authentication process based on the response
from the AVP component. If the response is “Allow/Allow with additional factor”, the
authentication server performs authentication by verification of password, and/or
asking the user to present additional information such as OTP.
5.3 Authentication Policy Specification

The security administrator specifies authentication policies in terms of provenance
abstractions. These policies will act a basis for detecting anomalies in the user login.
For instance, to prevent a password sharing scenario, the administrator may specify a
policy such as “If user’s previous login location does not match with the current login’s
location, and user previous login time is less than a predefined time” then it should be
treated as an anomaly because same user cannot be at different locations within the
predefined time.
5.4 Policy Engine

The policy engine comprises policy server, policy evaluation server, and policy data-
base. The policy server facilitates requests of the policy evaluation server, and policy
creation requests. The policy server facilitates the requests by talking to policy data-
base. The policy evaluation server serves authentication requests from the AVP, and
evaluates applicable authentication policies with the help of authentication provenance
to detect anomalies and respond to the AVP with the decisions.
5.5 Provenance Mechanism

The provenance mechanism facilitates authentication provenance collection, manipu-
lation, storage, and retrieval. It comprises two components: provenance server;
provenance store.
Provenance Server
When an authentication context is evaluated, the AVP makes a request to the prove-
nance mechanism with the authentication context information. The provenance server
collects data from the AVP and transforms it into meaningful form based on the
PROV-DM model. The provenance data—represented in RDF [19] form—will be
serialized into turtle form and stored in the provenance store. The policy server also
facilitates provenance requests from the policy engine and completes the requests by
querying the authentication provenance. We have used SPARQL [16] for querying the
authentication provenance.
Provenance Store
The provenance store maintains serialized provenance captured in RDF form. Cur-
rently, we are serializing the provenance onto text files and storing the files on the
provenance store.
6 Prototype Evaluation
In this section, we discuss the password sharing/stealing use cases evaluation.
6.1 Primary Detection

The PRAN prototype can prevent the brute force or dictionary attacks by using primary
detection mechanism. In this scenario, the webserver that hosts the application can
validate the user’s session with the provenance for detecting anomalies and deny the
login request. For instance, if a login request is coming from a system for which there
are repeated login failures that are greater than threshold then deny the login requests
from the system. This rule can be represented as below:
Denylogin $ provloginconsectivefailuresðcurrentmacÞ [ loginfailurethreshold

ð10Þ
The PRAN checks whether there are consecutive login failures for the system MAC
of the present login request. Usually, in case of brute force attacks, requests may come
from different systems or same system. To prevent anomalous login requests from
different systems, authentication policy rule can be changed to include other system or
contextual attributes.
6.2 Secondary Detection

The secondary detection method uses application level authentication policies, and
evaluates the rules to detect anomalies.
Use Case 1 - Alice and Bob Work in Different Locations—Bob Uses Alice
Credentials
The authentication policy will have a rule specified using provenance abstractions
“provlocmatch” and “provmacmatch”. The policy engine evaluates the rule “provloc-
match(current_location) and provmacmatch(currentmac)”
ð11Þ
where “current_location” and “current_mac” are the location, and system MAC of the
authentication context.
It may so happen that the user “Alice” carries her laptop and actually present in that
location and submitted the login request. In this scenario also, the system will allow
login because Alice system will match with the MAC used in the previous login.
Use case 2 - Alice and Bob Are Co-located—Bob Uses Alice’s Credentials
In this scenario, the PRAN prototype will deny the login request based on the
evaluation.
ð12Þ
Here, location matches with the previous login location. However, system MAC will
not match because Bob’s system MAC would be different from the Alice’s system
MAC.
Use Case 3 - Alice and Bob Are Co-located—Alice Uses Alice’s Credentials
This is a genuine scenario. However, the policy specifies “allow Alice to use Bob’s
system only if Alice’s system is crashed”.
ð13Þ
In this scenario, it checks that whether the user u logged into her system. In case of
system crash, the user u would not have logged into her system. Hence, the PRAN will
allow user login to the application using Bob’s system.
Use Case 4 - Bob Uses Alice’s Credentials on Alice’s System
This is a tricky scenario. The prototype cannot identify this scenario based on the
policy rules because Alice’s credentials are used on Alice’s system. However, the
administrator can specify an additional rule that if Alice has a habit of sharing cre-
dentials then inform the application to ask the user to provide additional information—
it is up to the application how it asks for additional information.
Use Case 5 - Eve Manages to Steal Alice’s Credentials
In this scenario, Eve manages to steal password by using shoulder surfing attack. Here,
we assume that Eve attempts to login to the application using a different system other
than Alice’s system. In this scenario, the rule the use case scenario 2 will detect the
anomaly.
ð14Þ
If current location matches with the previous login location and current system
MAC matches with the previous system login’s MAC or if location mismatches and
MAC matches with the previous login’s MAC.
Use Case 6 - Alice Uses Her Personal System to Login to the Application
In this scenario, Alice declares that she will use her home system which will be
captured as provenance. The check below will allow Alice to login to the system
ð15Þ
Where sa is a system attribute that identifies Alice home system.
6.3 Evaluation Results

We summarize the evaluation results in Table 1.
Table 1. Use case evaluation results.

Use case No Anomaly Result
1 Yes Detected—response “Deny”
3 No Response “Allow”
4 Yes Undetected—response “Allow”
6 No Response “Allow”
7 Limitations and Challenges
Even though, the PRAN prototype is promising in detecting anomalous login events, it
has a few limitations and challenges. Since it is based on provenance of different
entities, availability of all required attributes is critical—for instance, determining
location is possible only if, the user is willing to share location. The prototype needs
administrators to be knowledgeable about password sharing/stealing scenarios, and
provenance abstractions, to be able to create authentication policies for such scenarios.
There are some challenges to overcome in order to make the prototype a real-time
implementation:
• Collection—making provenance collection trustworthy, accountable, and seamless.
• Maintenance—dealing with performance overhead associated with processing,
retrieving and storing humongous provenance.
• Security—ensuring provenance security to prevent unauthorized modification.
8 Conclusions and Future Directions
We have presented the PRAN model, and the prototype. We evaluated the prototype
with use case scenarios, and provided its limitations and challenges. We observe that
this work is certainly promising to be a useful attempt in the area of strengthening
authentication leveraging provenance. We also observe that provenance abstraction
would need to be generated for all the scenarios utilizing authentication provenance
causal relationships to enable the administrators create policies. Even though, we
proposed the PRAN model for password based authentication, it can be used for other
types of authentications also. Our recommendations for future research are: utilizing
provenance as an authentication factor which can strengthen single-factor authentica-
tion, and can be unobtrusive, and convenient; dealing with performance overhead
associated with querying, processing of humongous provenance:is there any way, a
system can remove noise (unwanted provenance) from the collected provenance; uti-
lizing provenance for other security aspects with showing some advancement in
usability, convenience, scalable aspects.
References
1. Groth, P., Moreau, L.: PROV-overview. An overview of the PROV family of documents
(2013)
2. Braun, U., Shinnar, A.: A security model for provenance (2006)
3. Hasan, R., Sion, R., Winslett, M.: The case of the fake Picasso: preventing history forgery
with secure provenance. In: FAST, vol. 9, pp. 1–14 (2009)
4. Syalim, A., Nishide, T., Sakurai, K.: Preserving integrity and confidentiality of a directed
acyclic graph model of provenance. In: IFIP Annual Conference on Data and Applications
Security and Privacy, pp. 311–318. Springer, Heidelberg (2010)
5. Park, J., Nguyen, D., Sandhu, R.: A provenance-based access control model. In: 2012 Tenth
Annual International Conference on Privacy, Security and Trust (PST), pp. 137–144. IEEE
(2012)
6. Nguyen, D., Park, J., Sandhu, R.: Dependency path patterns as the foundation of access
control in provenance-aware systems. In: TaPP (2012)
7. Sun, L., Park, J., Sandhu, R.: Engineering access control policies for provenance-aware
systems. In: Proceedings of the Third ACM Conference on Data and Application Security
and Privacy, pp. 285–292. ACM (2013)
8. Nguyen, D., Park, J., Sandhu, R.: A provenance-based access control model for dynamic
separation of duties. In: 2013 Eleventh Annual International Conference on Privacy, Security
and Trust (PST), pp. 247–256. IEEE (2013)
9. Sun, L., Park, J., Nguyen, D., Sandhu, R.: A provenance-aware access control framework
with typed provenance. IEEE Trans. Dependable Secur. Comput. 13(4), 411–423 (2016)
10. Khan, H.Z.U., Zahid, H.: Comparative study of authentication techniques. Int. J. Video
Image Process. Netw. Secur. (IJVIPNS) 10(4), 9–13 (2010)
11. Bonneau, J., Herley, C., Van Oorschot, P.C., Stajano, F.: The quest to replace passwords: a
framework for comparative evaluation of web authentication schemes. In: 2012 IEEE
Symposium on Security and Privacy, pp. 553–567. IEEE (2012)
12. Cadenhead, T., Khadilkar, V., Kantarcioglu, M., Thuraisingham, B.: A language for
provenance access control. In: Proceedings of the First ACM Conference on Data and
Application Security and Privacy, pp. 133–144. ACM (2011)
13. Miles, S., Groth, P., Munroe, S., Moreau, L.: PrIMe: a methodology for developing
provenance-aware applications. ACM Trans. Softw. Eng. Methodol. (TOSEM) 20(3), 8
(2011)
14. Al-Karkhi, A., Al-Yasiri, A., Linge, N.: Discreet verification of user identity in pervasive
computing environments using a non-intrusive technique. Comput. Electr. Eng. 41, 102–114
(2015)
15. Yampolskiy, R.V., Govindaraju, V.: Behavioural biometrics: a survey and classification. Int.
J. Biom. 1(1), 81–113 (2008)
16. Prud, E., Seaborne, A.: SPARQL query language for RDF (2006)
17. Buriro, A.: Behavioral biometrics for smartphone user authentication. Ph.D. diss., University
of Trento (2017)
18. Extensible Access Control Markup Language (XACML), version 3.0. Oasis Standard, 22
January 2013
19. Carroll, J.J., Klyne, G.: Resource Description Framework ({RDF}): Concepts and Abstract
Syntax (2004)
20. Khan, W.Z., Aalsalem, M.Y., Xiang, Y.: A graphical password based system for small
mobile devices (2011). arXiv preprint: arXiv:1110.3844
21. Ruoti, S., Andersen, J., Seamons, K.E.: Strengthening password-based authentication. In:
Way@ Soups (2016)
22. Sun, H.-M., Chen, Y.-H., Lin, Y.-H.: oPass: a user authentication protocol resistant to
password stealing and password reuse attacks. IEEE Trans. Inf. Forensics Secur. 7(2),
651–663 (2012)
23. Reddy, R.S.C., Gopu, S.R.: A dynamic delegation model and prototype for an enterprise
context. In: 2017 IEEE Conference on Open Systems (ICOS), pp. 71–76. IEEE (2017)
The Deepest Vertex First Reboot:
Rebooting Network Edge Switches
in a Campus Network
Motoyuki Ohmori(B) , Satoshi Fujio, and Masayuki Higashino
Tottori University, Koyama-minami, Tottori 680-8550, Japan

{ohmori,s-fujio,higashino}@tottori-u.ac.jp
Abstract. Network edge switches require to be rebooted for some rea-

sons, e.g., updating a firmware, configuring a special behavior and so
on. This paper proposes the deepest vertex first reboot that can reboot
many network switches with less overhead downtime. The deepest ver-
tex first reboot can reduce downtime overhead by vertex contraction. The
deepest vertex first reboot then enables to reboot all switches by only
13-s overhead where the under-estimated ideal minimum downtime is
75 s in an actual campus network where there are more than 300 net-
work equipment.
1 Introduction
Even in the era of Software Defined Network (SDN) or Software Defined Infras-
tructure (SDI), network edge switches require to be rebooted in some situations,
e.g., updating a firmware, configuring a special behavior and so on. We, Tottori
University, replaced almost all network switches in our campus network with
newer network switches made by AlaxalA in September 2017. We are now facing
a serious problem that an edge switch requires a reboot to activate a new digi-
tal certificate for web authentication after installing the certificate. As a digital
certificate expires within few years or even few months, we have to periodically
install a new digital certificate into all edge switches, and reboot them. When an
edge switch is rebooted, an end host becomes disconnected, and this downtime
should be minimized.
In order to dig into the best order or timing to reboot an edge switch, this
paper proposes a method, the deepest vertex first reboot, to reboot edge switches.
This method represents a campus network as an unweighted directed rooted
spanning tree graph. This method then enables to obtain a set of edge switches
that can be simultaneously rebooted. This method then defines to reboot edge
switches in the deepest vertex first fashion, i.e., rebooting the farthest edge
switches from a core switch. This method can avoid unnecessary waiting time
for a switch reboot and then minimize the total downtime of a network.
Note that this paper focuses the case where almost all of edge switches require
reboots but non-reboot switches such as core, distribution and some edge switches
https://doi.org/10.1007/978-3-030-15035-8_14
152 M. Ohmori et al.
do not. It is out of the scope of this paper to reboot core and distribution
switches.
The rest of this paper is organized as follows. Section 2 defines preconditions
of this paper, and presents a campus network representation in this paper in a
graph theory fashion. Section 2 then proposes the rebooting method, the deepest
vertex first reboot. Section 3 explains a prototype implementation of the proposed
rebooting procedure including how to obtain a campus network graph. Section 4
then evaluates the proposed method using the prototype implementation in an
actual campus network. Section 5 finally concludes this paper.
2 Deepest Vertex First Reboot

2.1 Pre-condition
Suppose a campus network holds following pre-conditions.
1. non-reboot switches, i.e., core, distribution and some edge switches, do not
require to be rebooted,
2. an upstream edge switch may accommodate another downstream edge switch,
3. all switches are connected in the spanning tree, i.e., there is no alternative
path from an upstream edge switch to a downstream edge switch,
4. an edge switch cannot be scheduled to autonomously reboot,
5. an IP address of a core switch is given,
6. IP addresses of other switches are not, and
7. an IP address of a neighboring switch can be obtained in some way as
described in Sect. 3.
2.2 Campus Network in Graph Theory

This section presents notation of a campus network in a graph theory fashion in
this paper.
Figure 1 depicts an example of a campus network, and note that end hosts are
omitted. In Fig. 1, a campus network is represented as an unweighted directed
v0
v1 v2
v3 v4 v5 v6
v7 v8 v9
Fig. 1. An example of a representation of a campus network.

The Deepest Vertex First Reboot 153
rooted spanning tree graph G. V (G) and E(G) then denote a set of vertices and
arcs, respectively. A switch is a vertex, and i-th switch is denoted by vi ∈ V (G).
Let v0 be a vertex of a 0-th switch, i.e., a root. Let v0 be the only source vertex
with indeg(v0 ) = 0 in G, i.e., a core switch, while there are sink vertices vi with
outdeg(vi ) = 0 which are connected to no downstream switch. Downstream here
is a direction from the root v0 to other vertices. On the other hand, upstream is a
direction toward a root v0 . An upstream switch vi is then adjacent (i.e., physically
and directly connected) to another downstream switch vj by an unweighted arc
vi vj ∈ E(G) where i < j. On the other hand, suppose the downstream switch
vj has no arc back to the upstream switch vi for simplicity in discussions and
computational efficiency. Note that traffic may actually go through from vj to
vi in an actual network. Suppose non-reboot switches may not require to be
rebooted due to some reasons such that they accommodate no end host or do
not implement web authentication. Non-reboot switches depicted in Fig. 1 are v0 ,
v1 and v5 , and they are marked by prime ( ).
2.3 Vertex Contraction
As described in Sect. 2.1, some of switches, non-reboot switches, in a graph G

do not require reboots. In order to consider an order to reboot switches, we can
contract vertices of non-reboot switches. Note that a vertex contraction [3] here
is the same as an edge contraction in this paper since a graph G is a spanning
tree and two vertices to be contracted are always connected. Figure 2 depicts a
graph G/V (G) that contracts vertices of non-reboot switches from a graph G
depicted in Fig. 1. Note that V (G) is a set of vertices of non-reboot switches,
and “/V (G)” represents an operation of vertex contractions of all non-reboot
switches. In Fig. 2, v1 and v5 are contracted. A vertex contraction of a vertex vj
is an operation to:
1. remove an arc vi vj ,

2. union arcs to downstream vertices of vj to vi , and
3. delete vj .
Note that a root vertex v0 may also be able to be contracted in theory. A root
vertex v0 is, however, not contracted for simplicity in discussions in this paper.
Note that G/V (G) is then a spanning tree as well.
2.4 Depth of Vertex
A depth of a vertex is the number of arcs from a root in a graph, and a depth
of a root is 0. Let us focus on a depth in a contracted graph G/V (G), not in
a graph G. In Fig. 2, the depths of {v0 }, {v2 , v3 , v4 } and {v6 , v7 , v8 , v9 } are 0, 1
and 2, respectively. A depth of each vertex can be obtained as follows.
Let A be an adjacency matrix for G/V (G), and aij be an element (i, j) of A
where i and j denote indices of vi and vj , respectively. Let anij be an element (i, j)
of An . Note that indices of contracted vertices are not included in A and An .
v0
v3 v4 v2
v7 v8 v9 v6
Fig. 2. An example of vertex contractions.
Corollary 1. For an index of each vertex in G/V (G) except for a root v0 , there
exists one and only one n ∈ N such that an0j = 1, and n will be a depth of vj .
Proof. aij ∈ {0, 1} since G/V (G) is an unweighted graph. In accordance with
a nature of an adjacency matrix, if anij = 0 then n is the number of vertices
of a simple path from vi to vj and anij is the number of simple paths. Actually
anij ∈ {0, 1} because G/V (G) is an unweighted spanning tree, and there exists
only one simple path from vi to vj if exists. There also exists one and only one
simple path from v0 to vj . There, hence, exists one and only one n ∈ N such that
an0j = 1, and n is then a depth of a vertex vj .
Corollary 2. There exists m ∈ N such that Am = O and Am+1 = O. m is then

the maximum depth of vertices in a graph G/V (G).
Proof. A is an upper triangular matrix because G/V (G) is a directed graph and
there are no back arc to an upstream vertex. The main diagonal components aii
of A are all zeroes since an arc from vi to vi itself never exists. A is then a strictly
upper triangular matrix, and nilpotent, say with Ar = 0 [1]. A then holds {∃ m ∈
N | Am = O ∧ Am+1 = O}. The maximum depth of vertices is then m.
Corollary 3. Depths of all vertex can be obtained by computing An while n ≤ m.
Proof. An = O where n > m. Since G/V (G) is a spanning tree, there exists one
and only one simple path from v0 to vj . For each vertex in G/V (G) except for
a root v0 , there exists one and only one n such that an0j = 1 and n ≤ m. Depths
of all vertices can be then obtained by computing An while n ≤ m.
Let us take a look at an example of an adjacency matrix of G/V (G) depicted

in Fig. 2. The adjacency matrix A is expressed in Eq. 1. The square and cube of
A are Eqs. 2 and 3, respectively. The depth of {v2 , v3 , v4 } and {v6 , v7 , v8 , v9 } is
1 and 2, respectively. The maximum depth of vertices is then 2.
⎛ ⎞
a00 a02 a03 a04 a06 a07 a08 a09
⎜a20 a22 a23 a24 a26 a27 a28 a29 ⎟
⎜ ⎟
⎜a30 a32 a33 a34 a36 a37 a38 a39 ⎟
⎜ ⎟
⎜a40 a42 a43 a44 a46 a47 a48 a49 ⎟
A = ⎜
⎜a60
⎟
⎜ a62 a63 a64 a66 a67 a68 a69 ⎟
⎟
⎜a70 a72 a73 a74 a76 a77 a78 a79 ⎟
⎜ ⎟
⎝a80 a82 a83 a84 a86 a87 a88 a89 ⎠
a90 a92 a93 a94 a96 a97 a98 a99
⎛ ⎞
0 1 1 1 0 0 0 0
⎜0 0 0 0 1 1 1 1 ⎟
= ⎜
⎝0 0 0 0 0 0 0 0 ⎟
⎠ (1)
..
.
⎛ ⎞
0 0 0 0 1 1 1 1
A2 = ⎝ 0 0 0 0 0 0 0 0 ⎠ (2)
..
.
A3 = 0 (3)
2.5 Deepest Vertex First Reboot Procedure
In this paper, a downtime is a time required to reboot all switches requiring

reboots. To be more specific, a downtime is a time between:
• when the first rebooted switch stops forwarding packets and

• when the last rebooted switch becomes up and starts to forward packets.
Let us consider an ideal reboot operation that minimizes a downtime. The

ideal reboot operation is to reboot all edge switches requiring reboots at once.
This operation ideally minimize the downtime that is equal to max{vi ∈
G/V (G)|Treboot (vi )} where Treboot (·) denotes the time that the edge switch,
vi , reboots and becomes up. Note that this operation is ideal, and not feasible in
an actual environment because it is almost impossible for an operator to directly
connect to all switches via physical console cables, not a network. If an operator
accesses to all switches via a network, an operator may not be able to reboot
all switches at the same time because an order of arrivals of packets containing
reboot command will not be ensured.
To this end, we propose the deepest vertex first reboot to reboot the deepest
switches first. The deepest here means that the depth from a root v0 in a con-
tracted graph G/V (G) is the deepest among vertices of switches that are not
rebooted yet. Switches whose vertices are in the same depth are simultaneously
rebooted. Simultaneously here implies concurrently or parallely, and it depends
upon an implementation. The switches must be then confirmed if the switches

receive reboot commands, e.g., whether TCP acknowledgement for a segment
containing reboot commands is received, whether connections are closed or other
ways. The next deepest vertices are then examined to be rebooted. In this way,
the deepest vertex first reboot ensures that upstream switches are not rebooted
before downstream switches are not yet. This avoids the worst case where a
downstream switch waits to be rebooted after an upstream switch is rebooted
and becomes up, and then minimizes a downtime.
3 Implementation
3.1 Networking Utilities
In order to control a switch, we have implemented the Networking Utilities that
connect to a switch, and execute a command on a switch. The utilities are written
in Ruby, and can be run with Ruby 1.9.3 or later. The utilities require net/telnet
and net/ssh library, and then connect to a switch by telnet or SSH. The utilities
are free software licensed with 2-clause BSD license, and available on https://
github.com/ohmori7/netutils.
3.2 Different Graph per Campus

A campus network in a university can be represented as a single graph. It may,
however, incur a longer delay to wait for all switches at the same depth in a
single graph to start to reboot. We then divided a campus network per campus.
3.3 Automated Network Graph Production

In authors’ campus network, a network topology map and port list of switches
are not accurately maintained. It is then almost unknown which port of which
switch is connected to which port of which switch. In addition, the prototype
implementation focuses on the case where a switch should be installed a new
digital certificate and then rebooted. There is no document that clearly states
which switch runs web authentication. In other words, we do not know which
switches should be rebooted.
We then decided to automatically build the network topology map. To this
end, we employed either LLDP or CDP on all switches except for switches that
could not run both of LLDP and CDP. We also supported to statically define
such switches as a static neighbor of other switches.
As an IP address of a core switch in each campus is given, IP addresses of
neighboring switches are obtained by examining output of LLDP or CDP. All
other switches are then obtained by examining neighboring switches one by one.
In order to avoid a infinite loop, a switch is identified by a host name, and
already examined switches, i.e., upstream switches, are examined only once.
Note that telnet or SSH connections used for a network graph production
are kept for further procedures described in later sections.
3.4 Non-reboot Switch Detection
We here focuses on rebooting edge switches in a campus network to activate a

new digital certificate for web authentication. We then decided to regard network
equipment as being rebooted if and only if the following conditions hold:
1. network equipment is a switch,

2. web authentication is enabled, and
3. a switch is not rebooted after a new certificate is installed.
3.5 Rebooting Switches
After a network graph is produced and non-reboot switches are detected, vertices
of non-reboot switches are contracted. A depth of each vertex from a root is then
computed, all vertices of switches are grouped by depth. In order to minimize
a delay, all edge switches are logged in before rebooting. The switches of the
deepest vertices are then concurrently rebooted in the deepest vertex first fashion
using threads. The reason why reboots are concurrently not parallely is just
because of a Ruby implementation of threads that can run only one thread at
the same time. When a switch is rebooted, a switch is confirmed to accept a
reboot command if the switch holds one of the following conditions:
• CLI prompt is returned back to the Network Utilities, or

• a switch closes a TCP connection of a telnet or SSH connection.
After all switches of the deepest vertices are confirmed to start to reboot, the
next deepest vertices are examined.
4 Evaluations
4.1 Evaluation Environment
We have three main campuses, Koyama, Hamasaka, Yonago and tree branch
offices. Koyama and Yonago was connected by 10 Gbps. Koyama and Hamasaka
was connected by 1 Gbps by commercial VLAN service The branch offices were
connected to Koyama by 1 Gbps. The number of network equipment supported
by the Network Utilities in our campus network was 312 in total. 13 switches and
network equipment were then defined as a root. The number of edge switches
to be rebooted was 283, and some switches were not included because they
were shut down during evaluations due to building renovations. All switches
synchronized time using Network Time Protocol (NTP) [2] and its error would
be within 10 ms. All switches also run Rapid Spanning Tree Protocol (RSTP)
per VLAN in order to avoid a loop.
We implemented the Network Utilities on a server running on an ESXi as a
Virtual Machine (VM). We then implemented a code to measure time in Network
Utilities.
4.2 Time to Produce Network Graph

As described in Sect. 3.3, a network graph is automatically produced. If a network
graph is produced within a short time, the network graph production can be done
when rebooting all switches. If its production takes a long time, its production
should be done in advance. In order to clarify this issue, we measured time to
produce a network graph.
It took about 34.19 s in our campus network to automatically produce a
network graph. This time can be considered to be short enough to produce a
network graph at the same time when rebooting all switches. Our implementa-
tion then produces a network graph at the same time. Our implementation is,
however, limited to connected to 64 switches at the same time. This limitation
is intended to avoid memory and CPU consumptions that may incur a longer
delay resulting from swap out. It would be possible to increase this number or
remove this limit, and they are future work.
4.3 Overhead Waiting Time to Reboot Switches

The deepest vertex first reboot simultaneously reboots switches at the same depth
from a root. The deepest vertex first reboot then waits for all switches at the same
depth to surely receive a reboot command. This waiting time can be overhead
in comparison with the ideal reboot operation as described in Sect. 2.5.
We then measured this waiting time. As shown in Table 1, the maximum depth
was 4 while the actual maximum depth, i.e., without contraction, was 6. All
switches were rebooted within about 18.61 s in total. This waiting time was shorter
than the time to produce network graph, 34.19 s, because the telnet and SSH con-
nections were reused when rebooting switches. It can be then said that this waiting
time was shortened by connecting to all switches before rebooting them.
In addition, the time was interestingly not correlated to the number of
switches or depth from a root. For example, the number of 4 switches at the
depth 0 took 2.89 s while the number of 100 switches took only 2.98 s. Since their
time is similar, it may be said that the overhead time is correlated the maximum
time to send a reboot command. There were many models of switches, and the
difference of the models could be the cause.
It can be also said that vertex contraction reduces the total waiting time
because more than one second seems to require for a depth.
The detailed analyses will be future work.
4.4 Downtime
When switches are rebooted, downtime is the most important. We then measured
the downtime defined in Sect. 2.5 by logs of switches. A switch that was firstly
rebooted could log time when a reboot command was received. The time was
regarded as the beginning of downtime On the other hand, the time that a switch
lastly started to forward packets on the last VLAN was regarded as the end of
downtime. This time could be found by RSTP logs. The downtime was then 88 s.
Table 1. Overhead to reboot switches.
Depth Switches Time (sec.)

0 4 2.89
1 100 2.98
2 124 5.58
3 40 5.48
4 15 1.68
Total 283 18.61
Let us compare this downtime and the ideal minimum downtime. As described
in Sect. 2.5, the ideal minimum downtime is to max{vi ∈ G/V (G)|Treboot (vi )}
where Treboot (·) denotes the time that the edge switch, vi , reboots and becomes
up to start to forward packets. In order to find max{Treboot (vi )}, we computed
downtimes of several switches even though they were not all switches. AlaxalA
AX260A-08T, AX2230S-24P, AX2530S-08P, AX2530S-24T, AX2530S-24T4X,
AX2530S-48T and AX2530S-48T2X then took 63, 69, 74, 70, 68, 75 and 73 s to
start to forward packets, respectively. It could be said that max{Treboot (vi )} was
at least more than 75 s. Suppose that 75 s was max{Treboot (vi )}, and it could be
said that the overhead of the deepest vertex first reboot was about 13 s. It can be said
that the deepest vertex first reboot can reboot switches only by approximately 20%
downtime increase. Ones may wonder why 13 s are shorter than 18.61 s as shown
in Table 1. The cause would be that the switch requiring more time to reboot was
rebooted earlier than other switches at the same depth.
Note that investigating the actual max{Treboot (vi )} and more detailed anal-
yses are future work.
This paper has proposed the deepest vertex first reboot that can reboot many
network switches with less overhead downtime. This paper has tried to express
a campus network in a graph theory fashion. The deepest vertex first reboot can
then reduce downtime overhead by vertex contraction. The deepest vertex first
reboot then enables to reboot all switches by only 13-s overhead where the under-
estimated ideal minimum downtime is 75 s in an actual campus network where
there are more than 300 network equipment.
References
1. Hesselink, W.H.: A classification of the nilpotent triangular matrices. Compos.
Math. 55(1), 89–133 (1985)
2. Mills, D.: Network Time Protocol (Version 3) Specification, Implementation and
Analysis. RFC 1305 (Draft Standard). Obsoleted by RFC 5905, March 1992
3. Pemmaraju, S., Skiena, S.: Computational Discrete Mathematics: Combinatorics
and Graph Theory with Mathematica. Cambridge University Press, Cambridge
(2003)
Analyzing and Recognizing Pedestrian
Motion Using 3D Sensor Network
and Machine Learning
Ningping Sun1(&), Toru Tsuruoka2, Shunsuke Murakami2,

and Takuma Sakamoto2
1
Department of Human-Oriented Information System Engineering,
National Institute of Technology, Kumamoto College,
2659-2 Suya, Koshi, Kumamoto 861-1102, Japan
sningping@kumamoto-nct.ac.jp
2
Advanced Electronics and Information Systems Engineering Course,
National Institute of Technology, Kumamoto College,
2659-2 Suya, Koshi, Kumamoto 861-1102, Japan
Abstract. How to analyze and recognize pedestrian movements is an important

issue dependent on motion capture devices. In our work, we used two types of
popular 3D sensors such as 3D depth sensor and 3D motion sensor to construct a
sensor network for tacking motion of target because of their convenience and
low cost. In this paper, we first describe how to get data from the sensor network
and how to process raw data. Next, we provide algorithms for applying machine
learning to the analysis and recognition of human motions. Finally, we give
some evaluation experimental results.
1 Introduction
Research on human movement behavior has a long history and a wide range of
applications, especially in recent years, the research that reproduces human motion
behavior such as computer graphics, games, robots and artificial intelligence has made
remarkable achievements. Walking/running is one of the most characteristic chain
movements. Elucidation of the mechanism of walking and running is extensively
involved in training of athletes, and orthopedic surgery fields that specializes in the
treatment of lower limb disorders such as the foot, knee and hip joints. Many patients
with orthopedic diseases are causing obstacles by mechanical stress of actions centered
on walking. On the other hand, running is one of the most basic movements in sports
and there are various running forms. Each player has his/her own form.
At present, most doctors and coaches use their own naked eyes to observe the
walking/running style of patients or athletes, which lacks objectivity and appropri-
ateness. Therefore, we want to construct and develop a mobile pedestrian motion
analysis system by introducing advanced technology such as sensing and artificial
intelligence, which can provide a scientific analysis method based on accurate personal
data.

https://doi.org/10.1007/978-3-030-15035-8_15
162 N. Sun et al.
Advances in computer graphics for modeling human motion and the development
of sensing technology have led to the emergence of motion capture systems. From the
most simplified devices such as Kinect to large professional systems, motion capture
systems are capable of measuring and recording human motion. However, the problem
with conventional camera-based motion capture systems is that they can capture only a
narrow range of motion in front of the sensor and a large studio is required. They
cannot catch natural movements in a wide range of outdoor activities.
On the other hand, it is easier to convert human motion into digital data using
motion sensors such as acceleration sensor or angular velocity sensor. Although the
motion sensor devices can capture and record human motion, it is difficult to visually
recognize human motion behaviors by sensor only, for example, to determine whether
a person is walking or running by view the sensing data only.
Analysis of the motion data from the camera or the depth sensor is based on image
processing technology, while the analysis of the motion data captured by the accel-
eration sensor or angular velocity sensor is based on signal processing. Both have their
own strengths and weaknesses, by blending them we can make use of strengths and
complement the weaknesses. Recently the extraction and classification of sensor data
has also attracted people’s attention, and its importance has been understood year by
year. Apply artificial intelligence such as machine learning and deep learning to handle
large amounts of sensing data for detecting and recognizing human behavior is our
important subject.
In this study, we aim to develop a wearable mobile motion capture system by using a
depth sensor with a small amount of sensing data for modeling, and tracking human
movement using acceleration sensors and angular velocity sensors. We used data
acquired by the sensor network and machine learning to identify and recognize behavior
of pedestrian motion and tried to explore ways to improve the accuracy of recognition.
In this paper, we first describe how to get data from the sensor network and how to
process raw data. Next, we provide algorithms for applying machine learning to the
analysis and recognition of pedestrian motions. Finally, we give our experimental
results and discuss the possibility using our methods to analyze and recognize human
motion.
2 System Overview
We have been building a system in accordance with the following two points:
(1) collecting and storing pedestrian motion data through the sensor network, (2) using
machine learning to analyze motion data and identify pedestrian movements. Here we
give an overview of our system.
2.1 Collection of Motion Data Using Sensor Network

We collected the gait motion data using two kinds of motion sensors. First, in order to
identify person and analyze the human motion in a small range of the indoor space, we
constructed a pair of depth image sensor Kinects and obtained basic data of human
body skeleton information and motion postures.
Analyzing and Recognizing Pedestrian Motion Using 3D Sensor Network 163
When using a Kinect, the motion that can be taken is limited because when the
bone to be tracked enters the blind spot of the sensor, or when the body parts overlap,
the follow-up cannot be successfully followed. In order to overcome the deficiencies of
Kinect’s tracking of stereo motion and improve accuracy, we propose a method of
arranging multiple sensors to complement each other’s blind spots and perform motion
capture as shown in the left of Fig. 1. We combine and complement the obtained
position data so that it can cope with more complex motions [1].
TSND121
TSND121
TSND121
TSND121
TSND121
Fig. 1. (The left) Arranged Kinect pair, (the right) walking experiment with TSND121
Second, we use a group of small and communicable motion sensor TSND121s to

track and collect a wide range of walking/running data. As shown in the right of Fig. 1,
the sensors are attached to a total of five locations, left and right thighs, left and right
lower legs, and waist. When the movement of the body is measured, six groups of axis
inertial data are collected and sampled.
When measuring with motion sensors, characteristic errors often occur in accel-
eration and angular velocity, which affects accuracy. In our previous research, we
proposed a method to correct the motion sensor measurement error and derive it to the
uniaxial attitude angle [2, 3]. In this study, we have extended the method to six axes,
and implemented posture angle estimation using inertia information of six axis
excluding measurement error.
2.2 Analysis and Recognition of Pedestrian Motion Using Machine

Learning
As described above, the walking and running forms of a target usually are observed by
a doctor or a coach visually, and the judgment of whether the form is good or bad
depends on the experience and the degree of academic knowledge of the observer. In
order to correctly diagnose the patient’s changing symptoms and train athletes, it is
necessary to develop a scientific behavior analysis method based on accurate individual
gait measurement data.
Although motion sensor devices can detect human motion, sensors cannot recog-
nize and distinguish various complicated motion behaviors. To recognize chain motion
such as walking and running, an additional process of artificial intelligence is
164 N. Sun et al.
necessary. Therefore, we focus on applying artificial intelligence methods such as

machine learning and deep learning to deal with sensing data for human motion
recognition. In our work, we designed and implemented two methods for analyzing and
identifying pedestrian movements, one for the Kinect pair called Method 1 and the
other for the motion sensor TSND121 network called Method 2.
First, in Method 1, when performing motion capture using Kinect, we compensate
for blind spots by arranging a plurality of Kinects and obtain three-dimensional
coordinates of a set of joints for recognition by three-dimensional coordinate trans-
formation. To identify the postures of a human body, we employed two algorithms:
(1) to find similarities between training data and test data, (2) to construct a convo-
lutional neural network, and to identify postures by using machine learning of Sony
Neural Network Console. Section 4 details our approach.
Next, in Method 2, we used machine learning to identify the motion behavior with
the data of three-dimensional acceleration and angular velocity obtained by the motion
sensor network.
What matters very importantly in machine learning is to select features. Even if
there are many feature quantities, they cannot be identified with higher precision. It
may have very useful features, but it also contains some completely useless features. As
the number of features increases, the parameters learned by the machine learning model
increase, so that the calculation will take too much time. However, if the amount of
learning feature data is too small, the parameter estimation error will become large.
That is, the dimension of the feature vector is too small or too large to make a good
judgment [4]. In order to improve the accuracy of recognition and complexity, it is an
important subject for us to find an appropriate feature vector with the highest recog-
nition rate and the best calculation efficiency. We shall discuss this issue in Sect. 4 also.
3 Measurement and Sampling of Motion Data
Here we explain respectively how to obtain motion data from two kinds of sensors,
Kinect and TSND121, and how to sample the data for analysis.
3.1 Method 1
Kinect can detect 25 skeleton joints and track motions of up to six models at same time.
And each skeleton joint information includes such as jointType, three-dimensional
position coordinate, and tracking status (Tracked, Inferred, NotTracked), which are
updated at 30 fps to track motion. Horizontal angle is 70°, vertical angle is 60°, and
detectable distance is 0.5–4.5 m [1].
In our experiment we set a pair of Kinects to compensate blind spot as the left of
Fig. 1, we called them A and B. When shooting motion with two Kinects, the coor-
dinate set of the motion data obtained by each Kinect is represented on the coordinate
system A and B unique to each Kinect. It is necessary to unify the respective coordinate
systems before completing the coordinate of skeleton joint. According to the coordinate
system of a virtual monocular camera, we a apply series of affine transformations to the
coordinates set of each motion data [5].
Let DA be the 3D coordinate set of the motion data of Kinect A. When DA is

divided into frames, DA1 denotes the first frame before the affine transformation shown
in Fig. 2.
Fig. 2. DA1 before the affine transformation
Transformation Algorithm
Step 1. Translate DA to the coordinates of FootLeft of DA1 overlap the origin of the
virtual monocular camera.
Step 2. Find out the angle a about the y axis that the z component of the position
vector of FootRight becomes 0 and rotate DA around the y axis a degree.
Step 3. Find out the angle b about the z axis that the y component of the FootRight’s
position vector becomes 0 and rotate DA around the z axis b degree.
Step 4. Find out the angle c about the x axis that the z component of the position
vector of the Head becomes 0 and rotate DA around the x axis c degree.
Step 5. The final coordinates set D0A is obtained by (1), where v is the position vector
of the FootLeft of DA1 and R is the rotation matrix (2),
D0A ¼ RðDA vÞ ð1Þ

2 3
R11 R12 R13
R ¼ 4 R21 R22 R23 5 ð2Þ
R31 R32 R33
where R11 = cosacosb, R12 = −sinb, R13 = sinacosb, R21 = cosasinbcosc +

sinasinc, R22 = cosbcosc, R23 = sinasinbcosc − cosasinc, R31 = cosasinbsinc
− sinacosc, and R32 = −cosbsinc, R33 = sinasinbsinc + cosacosc.
As the result of the above process, the FootLeft is on the origin, the FootRight is on
the x axis, and the Head is in contact with the xy plane.
We give an example in Fig. 3, where the red model captured by Kinect A and the
blue model captured by Kinect B. In the images of A and B, the yellow skeleton joint
indicates that its tracking state is “inferred” that means the joint cannot be captured
correctly. Some “inferred” joints appear in red but not in blue, and vice versa. We
166 N. Sun et al.
proposed an algorithm that complements these hidden joints [1]. The complemented
result in the most right of Fig. 3 is used for our tracking and motion analysis.
A B A and B Complemented result
Fig. 3. The hidden joints complementation
3.2 Method 2
In each action experiment with motion sensor TSND121, we sampled three times, three
seconds each respectively, collected a total of 9000 data. Due to the efficiency of
Bluetooth communication between the server and the motion sensor, the amount of data
acquired was less than or greater than 9000 sometimes. In our experiments we sampled
data of walking and running, and each responding six-axis information included
acceleration (Ax , Ay ; Az ) and angular velocity (xx , xy ; xz ) acquired from the waist, left
thigh and left lower leg, right thigh and right lower leg like the right of Fig. 1.
The angle can be calculated by integrating the angular velocity with time. Usually
the measured value of the sensor contains an error. If we integrate as they are, the error
will accumulate, and the correct result will not be obtained. On the other hand, when
the angle is continuously obtained over a long period of time using the acceleration,
there is almost no error, but the error is large when they are obtained in a short time.
When angular angles are used to obtain angles continuously for a long period of time,
there are many errors, but the errors are small when calculated in a short time [6].
Therefore, we only extract the good part of both acceleration and angular velocity and
find the angles by complementing each other.
Let u be the roll angle of the sensor, h be the pitch angle, and t be the time.
Calculate ua and ha from the acceleration ðAx , Ay ; Az Þ by the (3) and (4).
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ua ¼ atan2 Ay ; A2x þ A2z ð3Þ
ha ¼ atan2ðAz ; Ax Þ ð4Þ
The motion angles are calculated using the complementary filter (5) and (6) where
R
xx dt is roll angle and R xy dt is the pitch angle obtained from angular velocity, ui is the
corrected roll angle and hi is the corrected pitch angle, where k = 0.95,
1 i n. Figure 4 gives the corrected results of walking experiment.
Z

ui ¼ k ui1 þ xx dt þ ð1 k Þua ð5Þ
Z

hi ¼ k hi1 þ xy dt þ ð1 kÞha ð6Þ
Fig. 4. Corrected roll (the left) and pitch (the right) angles of walking
4 Analyzing and Recognizing with Machine Learning
For the above Methods 1 and 2, we proposed some methods for analyzing and rec-
ognizing the motion behavior of the measured object using sensing data and machine
learning. In this section, we present our algorithms and approaches respectively.
4.1 Method 1
Recognition Using Posture Similarity. We chronologically vectorize the two pose
coordinate sets, one set of a and one set of b, and apply (7) to find the Euclidean
distance d between the joints of a and b. From our experiment results, we defined a
small threshold value e = 8. If the Euclidean distance d of the two poses is less than or
equal to the threshold e, the two poses are judged to be the same posture, otherwise it
will be judged as a different posture. If there are two or more tests in which the
Euclidean distance d for different motion is less than or equal to e, they are treated as
unrecognizable. Learning and recognition can be achieved by simply comparing the
test data set b with a previously prepared teacher pose model a. Accuracy can be
improved by preparing approximately 50 learning data and make the average of each
element as the model of the posture a.
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
X n
d¼ ðai bi Þ2 ð7Þ
i¼1
Recognition of Posture Using Convolution Neural Network. The convolution

neural network is a type of machine learning, and its feature vector is automatically
extracted by filtering an image using a convolution operation. In this study, we found
168 N. Sun et al.
that the nature of the posture can be represented by images, and the image conversion
operation does not affect the learning effect. We used Sony Neural Network Console
(SNNC) for construction, learning and evaluation of the convolution neural network.
First, it is necessary to convert the vector of the posture into a color image data. We
translated and scaled values of all the elements of the vector, so that values fall within
the range of 0 to 255, then the coordinate (x, y, z) are converted into one pixel as the (r,
g, b) pixel values. The learning process consists of five steps: (1) Input data of learning
and evaluation to the neural network. In this study, 50 images were prepared for each
posture as training data for learning, and the other 25 data were input for evaluation.
(2) Convolution: the size of the filter is 6 3, and the initial weight value of the filter
pffiffiffiffiffiffiffiffi
is w ¼ randð 2=nÞ devised by Kaiming He. This initial value was devised as an
optimum value when ReLU to be described later is used. (3) ReLU: add ReLU (ramp
x; x 0
function) as network activation function that is Rð xÞ :¼ . (4) Affine: the input
0; x\0
data is converted into a sequence of 100 elements by affine transformation. (5) Soft-
maxCrossEntropy: the data is classified using the Softmax function and cross entropy,
and the results are output.
4.2 Method 2
Feature Values Definition. Generally, pedestrians walk or run forward, so we only
use the acceleration Az from the acceleration (Ax ; Ay ; Az ) for analysis shown in red
curves of Fig. 5. In addition, the angular velocity xy plays an important role in all
pedestrian movement. In walking, stepping, and running, angular velocity xy repre-
sents the upward and downward angular velocity in the vertical direction shown in blue
curves of Fig. 5. We defined the standard deviation, maximum, and minimum of the
acceleration Az and angular velocity xy as the feature value in our analysis, which is
abbreviated as SDMM that is a 6-dimensional vector [3].
Feature Values Calculation. The peak waveform of Az and xy in Fig. 5 tell us that a
person can complete a step in about 700 ms. Therefore, we separate the sampling
interval by 700 ms. We further divided each part into two and made two windows of 0
to 350 ms and 350 to 700 ms. In doing so, the portion of the right thigh that is raised in
one step and the portion where the leg is lowered can be separated as a separate
window. We compute the SDMM for each of the two split windows and combine the
two partial SDMMs into the final feature SDMM of the 12-dimensional vector, which is
a row of the training data set or test data set. Next, we move the window from the
previous window 50 ms to the next 700 ms in order. Again, we calculate SDMM for
this new window. An integer n indicates the total number of rows in the training data
set or test data set in (8). In our experiments, when the training data was acquired for
3 s, there will be approximately 46 rows of SDMM for 3000 sets of acceleration z and
angular velocity y, with 12 columns per row, so up to 552 vectors should be calculated.
n ¼ ððlengthðdataÞ lengthðwindowÞÞ=50 ð8Þ

Fig. 5. Samples of walking (the left) and running (the right) motion data
Motion Recognition. We use k-NN for motion recognition. k-NN is a simplified

nonparametric machine learning method that is important and beneficial for data sets
that have noise or have a lot of variability. k-NN is used for classification, the input
consists of the k closest training examples, and the output is class member-
ship. A positive integer typically small, k, refers to the number of neighbors. Prediction
is based on the k neighborhood of test data. Like a general way, we set k = 3.
We prepared the training dataset and test dataset and made predictions to test what
the example was. Using (9) we calculate Euclidean distance d between the test dataset
and each training dataset, where i indicates the ith row of the dataset, 1 i n, and
j indicates the jth column of dataset, 1 j 12. Euclidean distance d provides a
distance between the feature SDMM of the given test motion and the SDMM of the
training motions. We rank the distances according to the minimum distance. Finally
estimating labels of test data, we use k-NN to majority vote its k immediate neighbors.
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
X12
di ¼ j¼1
ðtraining:SDMMi;j test:SDMMj Þ2 ð9Þ
4.3 Evaluation Experiment Results of Recognition

Here we give some evaluation experiment results of the analysis and recognition using
Method 1 and Method 2.
Evaluation Results of Method 1. We verified two methods of 4.1 using three types of
movement: step, raise right hand, and bow. For each movement we prepared 50 sets of
training data and 25 sets of test data individually. The average of 50 training data in
each movement was made as the teacher model a of (7) and threshold e = 8. Tables 1
170 N. Sun et al.
and 2 give the recognition results. F-measure of similarity is 0.88, SNNC is 1, which
tell us that converting the coordinate (x, y, z) of the postures into the color (r, g, b) of
the images make it becomes possible to use good efficient SNNC to recognize the
motion.
Table 1. Recognition of movement using posture similarity (e = 8)

Movement Step Raise right hand Bow Unknown Recall
Step 17 0 0 8 0.68
Raise right hand 0 20 0 5 0.8
Bow 0 0 22 3 0.88
Precision 1 1 1
Compared to the method of calculating the similarity by Euclidean distance, we

found that the method using the convolution neural network has higher movement
recognition accuracy. Also, we found that SNNC can recognize even if various data
with different angle and movement magnitude are given to the same type of posture.
One problem with the two methods is that the same motion cannot be recognized if the
speed is different. It is possible to be solved by changing the data extraction and
sampling method in our future work.
Table 2. Recognition of movement using Sony Neural Network Console (SNNC)

Movement Step Raise right hand Bow Unknown Recall
Step 25 0 0 0 1
Raise right hand 0 25 0 0 1
Bow 0 0 25 0 1
Precision 1 1 1
Evaluation Results of Method 2. We verified the method of 4.2 using three types of
movement: step, walk, and run. For each movement we prepared 9000 training data,
and one set of test data in which the first half is the data obtained from walk and the
second half from run respectively. We want to find out how close the test action type is
to already the given training types with (9) and k-NN classifier.
Figure 6 shows that recognition method of 4.2 has better recognition rate. Com-
pared to Method 1, Method 2 is simplified and easy-understanding. However, for
different measurement objects, the size of the sizing window used to divide the
waveform data samples affects the recognition accuracy.
As a solution, trial and error is performed to increase the learning data and adjust
the sizing window size of the raw data and prepare appropriate learning features set
before recognition.
Fig. 6. Recognition results of walking and running with (9) and k-NN
5 Conclusion
In this study, we proposed two methods, Method 1 and Method 2, to construct a

wearable mobile motion capture system by using a pair of depth image sensors and a
motion sensor network consisting of acceleration sensors and angular velocity sensors.
We applied machine learning to the sensing data for recognition of human motion
behavior and tried to explore ways to improve the accuracy of recognition.
From the results of our evaluation experiments, the methods of data processing and
recognition algorithms are feasible. Future tasks of this study include increasing the
amount of learning data, improving recognition accuracy, and expanding the scope of
recognition.
References
1. Sun, N., Murakami, S.: Human motion modeling from complementary skeleton joints of
multiple kinects. In: Proceedings of the 2018 International Conference on Mathematics,
Modelling, Simulation and Algorithms. Advances in Intelligent Systems Research, vol. 159,
pp. 131–135 (2018)
2. Sun, N., Sakai, Y.: New approaches to human gait simulation using motion sensors. In: 31st
International Conference on Advanced Information Networking and Applications Workshops
(WAINA) (2017). 16901760
3. Sun, N., Tsuruoka, T.: Pedestrian action recognition using motion sensor and k-nn classifier.
In: Proceedings of 2nd International Conference on Artificial Intelligence: Technologies and
Applications. Advances in Intelligent Systems Research, vol. 146, pp. 1–4 (2018)
4. Miyajima, S., Tanaka, T., Miyata, N., Tada, M., Mochimaru, M., Izumi, H.: Feature selection
for work recognition and working motion measurement. J. Robot. Mechatron. 30(5), 706–716
(2018)
5. Foley, J.D., Cardelli, L., van Dam, A., Feiner, S.K., Hughess, J.F.: Computer Graphics:
Principles and Practice, 3rd edn., pp. 263–286, 299–320. Addison-Wesley, New York (2015)
6. Lambrecht, S., Nogueira, S.L., Bortole, M., Siqueira, A.A.G., Terra, M.H., Rocon, E., Pons, J.
L.: Inertial sensor error reduction through calibration and sensor fusion. Sensors 16(2), 1–16
(2016)
Novel Interestingness Measures
for Mining Significant Association Rules
from Imbalanced Data
Safa Abdellatif1(B) , Mohamed Ali Ben Hassine1 , and Sadok Ben Yahia1,2
1
Faculty of Sciences of Tunis, University of Tunis El Manar,
LIPAH-LR11ES14, El Manar, 2092 Tunis, Tunisia
{Safa.abdellatif,mohamedali.benhassine}@fst.utm.tn
2
Department of Software Science, Tallinn University of Technology,
Akadeemia tee 15a, 12618 Tallinn, Estonia
sadok.ben@taltech.ee
Abstract. Associative classification is a rule-based approach that joins

Association Rule Mining and Classification to build classifiers that pre-
dict class labels for new data. Associative classifiers may generate an
overwhelming number of rules which are hard to handle. Delving through
these rules to identify the most interesting ones is a challenging task. To
overcome this problem, several measures have been proposed. However,
for imbalanced datasets, existing measures are no more reliable. In fact,
they tend either to favour rules of major classes and consider others as
uninteresting or only emphasize on the rules of minor classes and omit
other ones. In this respect, we propose five new measures which tend to be
fair for both types of classes regardless of their imbalanced distribution.
Extensive carried out experiments on real-world datasets show that the
new measures are able to efficiently extract significant knowledge from
minor classes without decreasing the global predictive accuracy.
Keywords: Association rules · Interestingness measures ·

Imbalanced datasets
1 Introduction
In the last decade, with the huge advance of technology and fast communication,
enterprises are daily collecting and storing large amounts of data from various
sources. In order to handle such overwhelming data by discovering unknown and
potentially useful information, several data mining techniques have been shown
to be of benefit. One of the most popular data mining techniques is Association
Rule Mining (ARM). It aims to detect relevant and valuable associations between
items in a database.
Associative classification (AC) is the integration of ARM and classification
for the purpose of constructing classifiers to predict the class labels of the new
data. Roughly speaking, ACs use predictive association rules which have as a
https://doi.org/10.1007/978-3-030-15035-8_16
Novel Interestingness Measures for Mining Significant Association Rules 173
consequence a pre-specified column called the class label, in order to be able to

infer that a new record belongs to a certain class. However, one of the problems
with the existing association rules algorithms is the overwhelming number of the
association rules. Hence, decision makers are unable to determine the interesting
and concise rules for prediction and consequently unable to make good decisions.
To overcome this problem, efficient evaluation of association rules has become
a compelling need rather than being a rational choice. To help carry out the task,
several interestingness measures have been proposed in the literature to filter and
sort rules with respect to given goals. However, in case of imbalanced datasets
where the class of interest (positive or minor class) is relatively rare as compared
to the other class (negative or major class), most of the proposed measures are
not efficient. In fact, they tend either to favour rules of the major classes whereas
considering ones of small classes as uninteresting and noisy. Alternatively, they
can only emphasize on the rules of minor classes without considering ones of the
major classes.
In order to overcome this shortcoming and be able to keep highly interesting
rules from both classes, we propose some new interestingness measures, which
aim to overcome problems of existing ones by tending to offer a high recall
on the minority class while maintaining a high precision of the majority class.
Experiments against six real-world datasets have been carried out with reference
to four assessment measures in order to evaluate the predictive power of the
proposed interestingness measures. Even though the classification is done using
one association rule-based classifier, the results can be easily generalized to other
classifiers. The remainder of this paper is organized as follows. Section 2 recalls
the basic concepts related to associative classification, Interestingness measures
as well as imbalanced datasets. The proposed approach is thoroughly described
in Sect. 3. Experimental results are shown in Sect. 4. Conclusion and perspectives
are sketched in Sect. 5.
2 Background and Related Work
In the following, we start by reviewing the associative classification as well as

interestingness measures.
2.1 Associative Classification and Interestingness Measures
Associative Classification (AC) is a rule-based approach proposed to classify

by first discovering a complete set of Class Association Rules (CARs) from the
training set and then using it to predict class labels for objects with an unknown
class. One of the problems of AC is that they have the ability to generate a huge
number of rules including a lot of uninteresting ones. To solve this problem,
it’s crucial to establish a “well-established” Interestingness Measure (IM) to
evaluate the quality and importance of an AR. Several IMs have been proposed
in order to find important ARs. These measures can be classified into three main
categories: objective, subjective and semantics-based measures. For simplicity, our
174 S. Abdellatif et al.
work only focuses on objective measures. There is a large number of objective

interesting measures available in the literature. One of the mostly used IM is the
confidence which is the conditional probability of having C given A and defined
as Conf idence(A → C) = P P(A∩C)(A) . Conviction is another IM which measures
the deviation from the independence by considering outside negation. Conviction
is defined as Conviction(A → C) = P (A)P (C)
P (AC)
. Laplace is also a measure used
for classification. It is a confidence estimator that takes the support in order
to calculate the Laplace value. As the value of the support decreases, the value
of Laplace also decreases. Laplace is defined as Laplace(A → C) = NN(AC)+1 (A)+2 .
Lift measures the dependency between A and C. It refers to the ratio of rules’
confidence to the probability of occurrence of the consequent. This ratio reflects
the positive or negative correlation of antecedent and consequence of the rule
P (AC)
and defined as Lif t(A → C) = P (A)P (C) .
To analyze this large number of measures, some properties have been pro-
posed for these measures. We consider in this section two sets of properties for
the IM related to a rule in the form of A → C. Piatetsky-Shapiro [9] proposed
three main properties which is desirable to be fulfilled by any IM.
– P1 : IM = 0 if A and C are independent, i.e., P (AC) = P (A)P (C). It means

that if A and C are independent in an AR then the measure should show
the least interest which is zero (or one for the lif t measure). A value greater
(less) than the least interest value indicates a positive (negative) correlation.
– P2 : IM monotonically increases with P (AC) when P (A) and P (C) remains
the same. This property states that the greater the support for AC, the
more interesting, the rule becomes. This means that the rule becomes more
interesting as far as the positive correlation between A and C increases.
– P3 : IM monotonically decreases with P (A) (or P (C)) when P (AC) and P (C)
(or P (A)) remain the same. P3 states that the rule becomes more interesting
if P (A) (or P (C)) gets smaller and P (AC) and P (A) (or P (C)) remain the
same.
Lenca et al. [5] have proposed five properties for the IM evaluation: (i) L1 . IM
is constant if there is no counterexample to the rule. L1 states that if P (AC) = 0
then IM’s value is constant. In other words, the rule with a value of confidence
equal to one should have the same interestingness value regardless the support
which contracts another property of Major and Mangano [7] that states that a
measure should combine support and association aspects; (ii ) L2 . IM decreases
with P (AC) in a linear, concave or convex fashion around 0+ . As a few numbers
of counterexamples are added, the L2 property describes in which manner the
IM decreases.
L3 . IM increases as the total number of records increases. L3 states that the
value of IM increases with the total number of records; (iii) L4 . The threshold
is easy to set. L4 states that it is easy to choose a threshold that separates
interesting from uninteresting rules; (iv ) L5 . The semantics of the IM is easy
to express: This property states that it is desirable that the semantics of the
measure is understandable. Several other factors influencing the interestingness

of a rule should be involved but somehow ignored in the literature. These factors
include the imbalance of class distribution and the cost of misclassification. In
the following, we will focus on the imbalance of class distribution factor.
2.2 Imbalanced Dataset Classification

Dealing with class imbalance has become a popular problem faced in data min-
ing. Imbalanced class distribution in datasets occurs whenever the number of
records of one class (called the major or negative class) is highly surpassing
the number of records belonging to the other class (called the minor or positive
class) [4]. Ironically, this class (the minor one) is often the one of more interest
focus and importance and which could be correctly recognized. The problem
of imbalanced datasets classification is encountered in several domains ranging
from sentiment analysis, medical diagnosis, etc.
However, classical classifiers are not able to handle the problem of imbalanced
datasets since they assume a balanced class distribution and do not take into
account the minority class. In particular, a classical associative classifier can
easily discover rules predicting the majority class. However, it is actually difficult
to discover rules predicting minor class since they consider them as uninteresting
rules and this is based on the same specific IMs which are not suitable in case
of imbalanced datasets.
One of these IMs is the confidence. In fact, it only considers the conditional
probability of rules. For example, the confidence of the rule A → C is 90%,
in the case that N (A) = 10, N (C) = 9000, N (AC) = 9 and N = 10, 000
where N is the size of data and N (X) denotes the frequency of X. Although the
confidence is very high, A and C are independent since P (AC) = P (A)P (C).
For this reason, although the rule A → C has very high confidence, it is not
considered as interesting since the value of confidence is exactly the probability
of C regardless of A. Finally, this rule might just be originated from noise.
We can also prove that Laplace is not the perfect measure in case of imbal-
anced datasets since it strongly relies on the support. In fact, as far as the value
of the support decreases, the value of Laplace also decreases which produces bad
results for rules of minor classes. Moreover, the conviction measure is not so help-
ful for the extraction of rules from minor classes. In fact, as mentioned above, the
conviction measure is defined as conviction(A → C) = Sup(A)Sup(C)Sup(AC)
. An associ-
ation rule is considered as interesting whenever its conviction highly surpasses 1.
However, if we take the example of a binary imbalanced dataset with C1 is the
major class and C2 is the minor class, the conviction value for the class C2 is writ-
ten as conviction(A → C2 ) = Sup(A)Sup(C
Sup(AC )
2)
. Since, sup(C1 ) 1 and C2 = C1 ,
2
then we have sup(AC1 ) sup(A) and conviction(A → C2 ) = Sup(A)Sup(C

Sup(AC1 )
1)
1.
To sum up, the conviction measure will not be able to extract important rules
from minor classes.
On the other hand, we can also prove that the lift measure favours rules
from minor classes. An association rule is considered as important if the lift value
highly surpasses 1. For major classes C1 , the lift value is defined as lif t(A → C1 )
sup(AC1 )
= sup(A)sup(C 1)
. Since sup(C1 ) 1 and sup(AC1 ) sup(A) then lif t(A →
C1 ) 1.
As a result, the lift measure may hardly find any interesting rule from major
classes but it is beneficial for the extraction of rules from minor classes. However,
one of the major drawbacks of the lift measure is its symmetry. In fact, the lift
measure takes A and C in equivalent position i.e. lif t(A → C) and lif t(C → A)
are the same, which is not always true in case of imbalanced datasets.
3 Proposed Interestingness Measures
Several interestingness measures have been proposed in the literature in order

to manage the overwhelming number of rules usually generated from real-world
datasets. Most of the IMs are not beneficial in case of imbalanced datasets since
they tend either to favour rules of major classes such as the confidence, Laplace
and the conviction measure or the rules of minor classes such as the lift measure.
We introduce, in the following, some new IMs which aim to overcome problems
of the existing ones by tending to be fair for both classes regardless of their
imbalanced distribution.
3.1 Description of Proposed Measures
We propose several measures which aim to capture rules which are beneficial for
major and minor classes. The first proposed IM is called M odif iedLif t and is
defined as follows:
lif t(A → C) P (A) P (AC)
M odif iedLif t(A → C) = ∗ = (1)
lif t(A → C) P (A) P (AC)
The first part of the formula refers to the ratio of the lif t measure of A → C
divided by the lif t measure of A → C. As the value of lif t(A → C) increases, A
and C are dependent. Consequently, there is a strong possibility that A and C are
dependent too. Moreover, the higher the lif t(A → C) is, the worse the rule A →
C. For these reasons, we have proposed lif t(A → C) as nominator and lif t(A →
C) as denominator. However, this ratio tends to have very high values. So we
propose to make some corrections by multiplying this ratio by the ratio of the
no occurrence possibility of premise divided by the occurrence probability of the
premise. As a conclusion, we may notice that the M odif iedLif t is getting larger
when (no occurrence of A implies no occurrence of C) surpasses (occurrence of
A implies the occurrence of C). This measure could be beneficial in case of
imbalanced datasets. In fact, if we take the example of a binary imbalanced
dataset with C1 as the major class and C2 as the minor one. The value of
M odif iedLif t for the rule A → C1 is equal to:
P (AC1 ) P (AC2 ) P (C2 )

M odif iedLif t(A → C1 ) = = 1
P (AC1 ) P (AC2 ) P (C2 )
since for the minor class C2 , we have P (AC2 ) = P (C2 ) and P (AC2 ) = P (C2 ).
On the other hand, the value of M odif iedLif t for the rule A → C2 is equal to:
P (AC2 ) P (AC1 ) P (A)

M odif iedLif t(A → C2 ) = =
P (AC2 ) P (AC1 ) P (A)
Based on these approximations, we may notice that the value of M odif iedLif t
for rules extracted from major classes is around one. However, for rules extracted
minor classes, this value depends on both the occurrence and non occurrence of
the premise part. We may notice that for this type of classes and based on our
approximations, we have:
M odif iedLif t(A → C2 ) < 1 if P (A) > P (A)

M odif iedLif t(A → C2 ) > 1 if P (A) < P (A)
As a result, it is obvious that if we have low support of A we may have a large

value of M odif iedLif t. Having low support of A could be probably due to the
length of the premise. That is worth mentioning that having a long premise part
could be useful for our type of data since extracted rules will be more specific
and may unveil new knowledge.
The following proposed measures derives from the first one M odif iedLif t.
One of these measures is called DM1 and defined as follows:
1 P (AC)
DM1 (A → C) = M odif iedLif t(A → C) ∗ = (2)
P (AC) P (AC)P (AC)
It inversely depends on the value of P (AC) which is suitable in case of imbalanced

datasets since the value of P (AC) for minor classes is generally lower than the
value of P (AC) for major classes.
DM2 and DM3 also derive form M odif iedLif t and depend inversely on the
probability of occurrence of the conclusion P (C) which is also convenient for the
case of imbalanced datasets since P (C) of minor classes is much less than P (C)
of major classes. These two measures are defined as follows:
1 P (AC)
P (C) P (AC) P (C)
P (A) P (AC)P (A)

P (C) P (AC)P (C)
Last but not least, DM4 is also a derived measure from M odif iedLif t but
it inversly depends on the probability of occurrence of the premise P (A). This
could be suitable for imbalanced datasets since it favors rules having low P (A)
which generally correspond to long antecedents. DM4 is defined as follows:
1 P (AC)
P (A) P (AC) P (A)
3.2 Properties of Proposed Measures

Table 1 shows six different properties mentioned in Subsect. 2.2 for the five pro-
posed measures as well as for the existing measures also described in Subsect. 2.2.
“Yes” means that the proposed measure has that property and “No” means that
the proposed measure does not have that property.
Table 1. Properties of proposed interestingness measures
Measure Properties
P1 P2 P3 L1 L2 L3
ModifiedLift No Yes Yes No Yes Yes
DM1 No No Yes No Yes Yes
DM2 No Yes Yes No Yes Yes
Confidence No Yes Yes Yes Yes No
Laplace No Yes Yes No Yes Yes
Lift Yes Yes Yes No Yes No
Conviction Yes Yes Yes Yes Yes No
4 Experimental Results
4.1 Data Collection and Experimental Setup
In order to evaluate the efficiency of the newly proposed IMs, we have selected
six real-world datasets: five of them are taken from the UCI repository [8] and
the last one is composed of posts collected from Twitter. The characteristics
of these datasets are summarized in Table 2. For each dataset, the number of
instances, the number of attributes and the imbalance ratio are provided. The
latter is defined as the ratio of the number of instances of the major class divided
by the number of instances of the minor class.
4.2 Results of the Experiments

In the following, each of the five newly proposed measures is applied, instead of
the existing ones as the primary sorting criterion in the rule selection phase and
this using the rule-based approach ARCID [1,2]. The adapted ARCID with the
Table 2. Datasets used for the evaluation
Datasets # Instances (#Maj/#Min) # Attributes Imbalance ratio

Breast Cancer (BC) 286 (201/85) 10 2.36
Tic-tac-toe (TTT) 425 (276/149) 10 1.68
Mofn 1323 (1031/292) 11 3.53
Post-operator (PO) 88 (64/24) 9 2.66
German 1001 (701/300) 25 2.33
Tweets 259 (163/69) 25 1.68
five new measures will be compared to the adapted ARCID with four existing
measures (i.e. Confidence “Conf”, Laplace “Lap”, Lift and Conviction “Conv”)
and several well-known classification approaches including CBA [6], C4.5 [10]
and Ripper [3]. Experimentation will be carried with reference to four assessment
measures which are the Global Accuracy, Geometric Mean, Weighted Accuracy
and F-measure.
Global Accuracy. Global Accuracy is the proportion of instances which are

correctly classified. The experimental results are sketched in Table 3. As can
be seen from this table, the adapted versions of ARCID with the confidence,
Laplace and Conviction measures yield better results than ARCID with the novel
measures. This could be explained by the fact that these measures favour rules of
major classes while considering those belonging to minor classes as uninteresting.
Moreover, we may also notice that the state-of-art algorithms perform slightly
better than the ARCID versions adapted with the novel measures. This owes to
the fact that these algorithms aim to maximize the global prediction accuracy
and minimize the error rate to which minor class rarely contributes. However,
if we take a closer look to the average rate of the global accuracy, we may
notice that the adapted versions of ARCID with the novel measures offer good
results compared to others even though they emphasize on efficiently predict the
instances of minor classes.
Geometric Mean. As we mentioned above, the ARCID approach using the

newly proposed measures yields slightly lower results in terms of accuracy. How-
ever, in the case of imbalanced datasets, accuracy is not the perfect measure to
evaluate the efficiency of a classifier. Indeed, it puts more focus on major classes
than minor ones. For this reason, we propose to use an additional assessment
measure in order to assess the balance between the performances of both major
and minor classes this assessment measure is called the Geometric mean.
According to these results, we may conclude that the novel measures are
statistically proved to be suitable for the selection of rare but high-quality rules.
Table 3. Global accuracy results
Datasets Proposed measures Existing measures Existing algos

ML DM1 DM2 DM3 DM4 Conf Lap Lift Conv CBA C4.5 Ripper
BC 0.76 0.75 0.76 0.76 0.76 0.73 0.74 0.74 0.73 0.73 0.73 0.74
TTT 0.86 0.86 0.86 0.86 0.86 0.92 0.92 0.58 0.74 0.75 0.80 0.98
Mofn 0.78 0.56 0.65 0.88 0.60 0.93 0.90 0.23 0.81 0.96 0.85 0.90
PO 0.66 0.66 0.66 0.62 0.66 0.74 0.74 0.59 0.74 0.73 0.73 0.73
German 0.71 0.66 0.69 0.66 0.71 0.88 0.88 0.48 0.88 0.72 0.75 0.71
Tweets 0.78 0.73 0.75 0.72 0.75 0.65 0.65 0.58 0.78 0.71 0.67 0.65
Avg rank 4.50 7.00 5.50 6.16 5.16 4.00 3.66 11.00 5.00 6.00 6.66 4.83
Avg rate 0.76 0.70 0.73 0.75 0.72 0.80 0.80 0.53 0.78 0.76 0.75 0.78
Table 4. Geometric mean results

BC 0.7 0.7 0.7 0.69 0.7 0.38 0.19 0.67 0.42 0.58 0.47 0.59
TTT 0.89 0.89 0.89 0.89 0.89 0.93 0.93 0.51 0.74 0.58 0.68 0.97
Mofn 0.85 0.66 0.75 0.89 0.72 0.93 0.49 0.29 0.41 0.92 .65 0.85
PO 0.35 0.35 0.35 0.34 0.35 0 0 0.33 0 0 0 0
German 0.38 0.69 0.61 0.69 0.61 0 0 0.53 0 0.38 0.66 0.58
Tweets 0.77 0.74 0.76 0.73 0.76 0.25 0.25 0.6 0.72 0.68 0.62 0.65
Avg Rank 3.16 3.16 3.00 3.83 3.16 7.00 8.66 8.83 8.83 7.16 7.16 5.50
Avg Rate 0.65 0.65 0.67 0.70 0.67 0.41 0.31 0.48 0.38 0.52 0.51 0.60
Weighted Accuracy. According to the results depicted in Table 5, it can be

observed that the ARCID approach using the newly proposed measures provides
statistically better results. Based on these findings, we may assert that the new
measures successfully tend to focus on increasing the accuracy of the minor class
while trading off the accuracy of major class (Table 4).
F-measure. Table 6 presents the f-measure results obtained from the experi-
ments. It shows that the ARCID versions using the newly proposed measures
outperform all the rest terms of F-measure which owe to the fact that it asserts
an absence of a bias towards minor classes.
Considering all of the above, we may conclude that the ARCID approach
using the newly proposed measures gets better ranks than ARCID using the
existing measures as well as the state-of-art algorithms and this is according
to the four evaluation measures used. To sum up, we may conclude that the
newly proposed measures are beneficial in case of imbalanced datasets since they
efficiently extract significant knowledge from minor classes (proven using the
Geometric mean, the Weighted Accuracy and the F-measure) without decreasing
the predictive accuracy of the classifier (proven using Global Accuracy metric).
Table 5. Weighted accuracy results

BC 0.70 0.69 0.70 0.70 0.70 0.37 0.13 0.68 0.42 0.65 0.65 0.66
TTT 0.85 0.85 0.85 0.85 0.85 0.92 0.92 0.38 0.74 0.77 0.80 0.99
Mofn 0.76 0.68 0.70 0.83 0.66 0.93 0.39 0.25 0.29 0.98 0.85 0.86
PO 0.49 0.49 0.49 0.46 0.49 0 0 0.44 0 0 0 0
German 0.60 0.66 0.62 0.65 0.64 0 0 0.45 0 0.69 0.69 0.63
Tweets 0.77 0.73 0.74 0.71 0.74 0.82 0.82 0.66 0.77 0.66 0.72 0.64
Avg Rank 3.80 4.66 4.16 3.83 4.16 5.50 7.00 7.50 8.66 6.16 6.66 6.00
Avg Rate 0.69 0.68 0.68 0.70 0.68 0.50 0.37 0.47 0.37 0.62 0.61 0.63
Table 6. F-measure results

BC 0.57 0.56 0.57 0.56 0.57 0.22 0.07 0.52 0.27 0.44 0.33 0.45
TTT 0.82 0.82 0.82 0.82 0.82 0.91 0.93 0.51 0.64 0.49 0.61 0.97
Mofn 068 0.53 0.57 0.78 0.48 0.87 0.39 0.36 0.29 0.92 0.57 0.78
PO 0.18 0.18 0.18 0.17 0.18 0 0 0.15 0 0 0 0
German 0.23 0.55 0.46 0.55 0.47 0 0 0.31 0 0.25 0.53 0.43
Tweets 0.71 0.68 0.70 0.67 0.70 0.11 0.11 0.60 0.66 0.54 0.52 0.58
Avg Rank 3.50 4.40 3.16 3.66 3.50 7.33 8.66 8.00 9.00 7.50 7.50 4.33
Avg Rate 0.53 0.55 0.52 0.58 0.54 0.35 0.25 0.41 0.31 0.44 0.43 0.53
5 Conclusion
This paper deals with the problem of association rule selection for prediction
in case of imbalanced datasets. Existing measures do not yield good results in
case of imbalanced datasets since they tend either to favour rules of major class
and consider others as noise or emphasize only on the rules of minor classes and
omit others. This paper proposes five new IMs to keep highly interesting rules
from both types of classes simultaneously in order to be able to offer a high
recall on the minority class while maintaining a high precision of the majority
class. Experiments show that the newly proposed measures offer a better pre-
dictive power compared to the existing ones. In our future work, we intend to
study the predictive power of the IMs with several datasets in text mining and
sentiment analysis. Moreover, we plan to apply these new IMs on well-known
algorithms such as CBA [6] in order to make them suitable for the classification
of imbalanced datasets.
References
1. Abdellatif, S., Ben Hassine, M.A., Ben Yahia, S., Bouzeghoub, A.: ARCID: a new
approach to deal with imbalanced datasets classification. In: International Confer-
ence on Current Trends in Theory and Practice of Informatics. Springer (2018)
2. Abdellatif, S., Ben Yahia, S., Ben Hassine, M.A., Bouzeghoub, A.: Fuzzy aggre-
gation for rule selection in imbalanced datasets classification using choquet inte-
gral. In: 2018 IEEE International Conference on Fuzzy Systems, FUZZ-IEEE 2018,
Rio de Janeiro, Brazil, 8–13 July 2018 (2018)
3. Cohen, W.W.: Fast effective rule induction. In: Proceedings of the Twelfth Inter-
national Conference on Machine Learning, pp. 115–123 (1995)
4. Hu, B.G., Dong, W.M.: A study on cost behaviors of binary classification measures
in class-imbalanced problems. arXiv preprint arXiv:1403.7100 (2014)
5. Lenca, P., Vaillant, B., Meyer, P., Lallich, S.: Association rule interestingness mea-
sures: experimental and theoretical studies. In: Quality Measures in Data Mining,
pp. 51–76. Springer (2007)
6. Ma, Y., Hsu, W., Liu, B.: Integrating classification and association rule mining.
In: Proceedings of the Fourth International Conference on Knowledge Discovery
and Data Mining (1998)
7. Major, J.A., Mangano, J.J.: Selecting among rules induced from a hurricane
database. J. Intell. Inf. Syst. 4(1), 39–52 (1995)
8. Merz, C.: UCI repository of machine learning databases (1996). http://www.ics.
uci.edu/∼mlearn/MLRepository.html
9. Piatetsky-Shapiro, G.: Discovery, analysis, and presentation of strong rules. In:
Knowledge Discovery in Databases, pp. 229–238 (1991)
10. Quinlan, J.R.: C4.5: Programs for Machine Learning. Morgan Kaufmann,
San Francisco (1993)
Extended Scheme Mediation Integration Model
for Information Systems Project Proposal
William Chaves de Souza Carvalho(&), Pedro Frosi Rosa,

and Flávio de Oliveira Silva
Federal University of Uberlandia, Uberlândia, Brazil

{william.carvalho,pfrosi,flavio}@ufu.br
Abstract. The demand for semantic sharing between information systems is

steadily growing in virtually all business domain. In fact, these systems form the
basis of business management, lead economic and decision-making strategies,
and manage communication with partners. Interoperability has become a pri-
ority requirement to and in order to meet the needs of applications have
autonomy, scalability, transparency and extensibility characteristics which is
crucial for data conflicts not to hamper efforts to integrate and transparently
share information from a variety of sources. Research has shown that the
evaluation of data semantics is a promising approach to dealing with this
problem. This work presents the project of creating a model based on mediation
of schemas to facilitate the understanding of shared data and help to identify and
adapt data relevant to the user context.
1 Introduction
Information Systems (IS) face two major challenges in their operation. On an hand,
there is a need to ensure interoperability between the heterogeneous systems that are
developed, in general, independently based on a single domain representation. On the
other hand, the operational environment is dynamic which leads to continuous changes
in these systems. Thus, it is necessary to optimize the process of reusing the modules
between the different systems to accelerate the developments. In analyzing these
challenges, we find that the problem of semantic sharing is a major factor that degrades
interoperability and prevents reuse between ISs.
The possible way to deal with the problem of semantic heterogeneity is to reduce or
even eliminate terminological and conceptual incompatibilities. Therefore, establishing
a common understanding (terminological and conceptual), with multiple points of
view, can help establish a communication base between users, manage interoperability
between systems and improve the engineering process of reuse.
Over the centuries, the constant evolution of communication technologies has made
the sharing human impetus increasingly more important. Sharing information is not a
new idea and in this paper, we focus on some challenging interoperability issues that
emerged with the evolution of communication technologies.
Neither the sharing of information between people nor the exchange of data
between information systems are new ideas. Nevertheless, new challenges are created
by the improvement of interconnection technologies of computational agents and
https://doi.org/10.1007/978-3-030-15035-8_17
184 W. C. de Souza Carvalho et al.
improved algorithms for processing and retrieval of large volumes of data. The solu-
tions must meet different requirements and criteria to accomplish the user needs [1]:
autonomy of the systems, extensibility of the architecture to control the addition or
removal of IS, scalability of the system capacity faced with the increase on the users
amount and transparency regarding location and data format. The solutions are based
on the notion of interoperability, that is, the implementation of a such collaboration
between systems that allows data and service sharing to respond to information
requests. Resolving a query in an interoperable environment involves integrating dis-
tributed data across multiple heterogeneous systems raises some questions:
1. Which ISs may contain all or part of the desired information?
2. What data are relevant to the expressed needs?
3. How to transform, to adapt and to pair the data in order to obtain a consistent and
user-friendly result?
The information integration is preceded by resolution of three types of conflicts:
• Syntactic conflicts result from using different data models from one system to
another; beyond what, different concepts are used to structure the same information.
• Schematic conflicts arise from different structures and classifications of information.
This is closely related to the choices made during designing the project.
• Semantic conflicts arise from differences in the interpretation of information shared
by various application domains: nomenclature conflicts (taxonomic and linguistic
problems), value conflicts (unit problems, scales), and cognitive conflicts (meaning
problems).
Schematic and semantic conflicts are closely related. The meaning of information
must be clearly defined, and semantic conflicts resolved before schematic conflicts are
identified and addressed.
Many solutions have been proposed to consider these data conflicts, that guarantee
at least the autonomy, the composability and the resolution of syntactic conflicts.
Multi-base approaches rely on the use of a query language and a common repre-
sentation model [2]. Each SI exports its information as a schema described in the
common model (usually an object-oriented model), the multi-base language, extended
from SQL (Structure Query Language) or OQL (Ontology Query Language), allows a
multi-site query. The multi-base approach is extensible but does not offer localization
transparency. The resolution of schematic and semantic conflicts remains entirely on
user’s responsibility.
Federated approaches rely on integration. Each SI exports a schema in a common
model, so the various exported schemas are integrated into federated schemas. A fed-
erated schema allows access to shared data in a uniform and global way [3]. Schematic
and semantic conflicts are solved by the integration process. Transparency is ensured,
just by posing a query on the federated schema. However, extensibility and scalability
are criteria that are not adequately met; which indicates that federation is an approach to
be considered when integration involves a small number of information systems and
when these systems go through few evolutions.
Mediation approaches extend the federation by providing more flexibility.
Mediation rely on an essential component, named mediator, which is responsible for
Extended Scheme Mediation Integration Model 185
meeting the integration needs based on the knowledge made available [4, 5]. The
mediator finds the available information and solve the detected schematic and semantic
conflicts. A secondary component serves as interface to the ISs and solve syntactical
conflicts by providing data to the mediation model.
Two types of mediation can be distinguished according to how data conflicts are
resolved:
1. Mediation of schemes that previously builds a database of information about the
participating ISs which gives the mediator the means to do his integration work.
2. Context mediation in which the mediator uses semantic information to resolve
queries dynamically without prior knowledge of which information systems belong
to the context.
Schema mediation is a direct extension of the federated approach [1, 6, 7] with better
scalability and often better scalability (object interfaces, rules-based language) [8, 9].
Context mediation seeks to discover data semantically close. This such mediation can
find and adapt information to ensure total transparency [10–13, 21, 22, 27, 28].
This paper presents the main aspects of a under development extended scheme
mediation integration model for information systems project proposal. When finalized,
the model will be composed by a methodology, a model and an architecture to build a
knowledge base to integrate information corresponding to well-defined usage profiles.
The mediation model will make it possible to represent the data structure in a common
model, adapt and compose the information in order to ensure total transparency. This
model will also associate semantics with shared information to compare and identify
relevant data. An incremental integration mechanism guided by semantics will allow its
adaptation to dynamic and broad environments.
Section 2 provides an overview of solutions based on context mediation and then
introduces the key points of for a integration model based on extended scheme medi-
ation. Section 3 describes a typical example of a cooperation environment and explains
how our approach can solve problems related to interoperability of IS. Section 4 con-
cludes the article and presents research perspectives.
2 Related Work
Context-mediated solutions are based on models of knowledge representation capable

of describing, to a certain extent, the semantics conveyed by information and tools for
comparing and unifying the semantics of information independently of the inherent
structures. The notion of semantics of an entity can’t be represented in an absolute way.
An entity must be linked with other entities for a meaning to be associated with it. The
semantic description of an entity only makes sense in relation to a specific context.
A context can be defined as the finite set of concepts, relationships, constraints, and
rules that describe an application domain. Each solution defines contexts on which the
semantic aspect of shared information is based. The information associated with the
same context can then be easily classified and compared since they share the same
semantics: it is the phase of semantic approximation. This phase is more delicate when
it comes to reconcile two information respectively defined on two previously disjointed
contexts. It is then a question of reconciling the contexts to estimate if it is possible to

reconcile this information and to deduce the way to compose them.
A context can be represented by a conceptual diagram that describes a domain area
of application. It is composed of concepts related to each other (inheritance relations,
compositions relations, logical relations). A concept is usually defined from the content
of an ontology. The ontology [14] is the specification of a conceptualization, the
expression in a language of terms and concepts. An ontology can be represented using
a semantic network, a terminology graph, a conceptual graph, a set of logical rules, or
an object-oriented schema [4]. The content of an ontology remains close to a specific
field of application even if some projects are moving towards the definition of generic
ontologies [15].
Solutions are categorized into two groups depending on whether the mediation uses
a single ontology or inter-ontological relationships. In the first case, single ontology,
the contexts of the information are constructed from the same conceptual universe and
the contextual reconciliation passes by a direct comparison of the concepts which then
have a common basic semantic unit.
The Coin project [12] is based on the description of the semantics of the values of
the exchanged data. Each piece of data specifies the semantics of its values by indi-
cating, for example, the type of unit and its meaning. An ontology provides a
vocabulary to describe the meaning of these values. Each SI has a local context that
describes the semantics of values manipulated locally. A reference context describes
the domain of use of the values and the transformation rules making it possible to
reconcile the semantics of values. These rules use the frame logic to derive a translation
path from a source context to a target context and adapt the representation of the
information [6]. The user can query the system without worrying about the format of
the data presented to him in his local format. The identification of semantic conflicts is
restricted to differences in the interpretation of values. Localization and composability
remain the responsibility of the user.
On2broker [13] is particularly interested in extracting data from the content of web
pages. Interoperating here consists in extracting the relevant data from the content of the
web pages, discovering the links with information located on other pages, combining
this data. A semantic model adds information to the content of web pages in the form of
proprietary HTML tags in the manner of XML. Ontology is defined as object classes and
deduction rules that extend the F-logic language [16]. These are references to these
classes that are linked to the content of web pages. A research and analysis process
continually scam participating web sites to build a global fact base that will be used to
infer ontology rules about the facts. Ontology takes the appearance of a global schema to
access data. The construction of a fact base compromises the scalability of the system.
The Observer project [10, 17] is based on a hierarchy of ontology servers described
in terminology logic to define utilization domains. Each SI must define its own data in
relation to the concepts of one or more ontologies. Rewriting rules and transformation
functions define the local interpretation of information. A user asks a terminological
query on the concepts of an ontology. This request can be sent over the entire domain
and only servers whose context agrees with that of the request provide answers.
Semantic reconciliation is based on predefined 2-2 inter-ontology relations and
synonymic relations.
InfoSleuth [18] is based on an agent architecture. Resource agents store the

information provided by the data sources using the vocabulary of ontological agents.
Each resource must be registered to a broker agent to ensure the location of the
information. A user agent provides a query interface to users and manages a query
using broker functionality. Information discovery is supported by specialized agents
that extend the architecture to handle multiple ontologies simultaneously through
predefined inter-ontology relationships.
The use of ontologies to describe a universe of speech is the common point of many
solutions. Defining an ontology server as a semantic reference is nevertheless not
without problems:
1. Identifying the concepts of an ontology requires a consensus on the part of the
interveners.
2. Modeling of the ontology is a delicate step.
3. Semantic universe is limited by the content of the ontology.
Taking into account several ontology servers brings out the problem of matching
between heterogeneous representations [19].
3 Semantic Approach
Our proposal will rely on incremental integration of relevant information dynamically

discovered on information systems. It will take advantage of the robustness of the
schema mediation approach and will combines it with the semantic matching tech-
niques of context mediation. The integration work is simplified once the information
has a contextual description that guides their integration it:
1. Allows the understanding of shared data.
2. Supports contextual reconciliation for comparing contexts and identifying infor-
mation in semantic relation.
3. Ensures the location of the information.
4. Facilitates the generation of the integration information and adapts the data to a
reference context.
Our model will adopt a technique of contextual reconciliation independent of an
ontological language, based on the specification of local contexts. That way, infor-
mation identified on heterogeneous but overlapping contexts can be compared without
the definition of inter-ontology relationships. Contextual reconciliation will be based on
a technique of comparing the concepts that make up two contexts. So, the similarity of
concepts can be evaluated according to their taxonomic similarity, their intrinsic
properties, and their semantic neighborhood (that is, relationships between concepts).
We define a concept as an elementary semantic unit that could be represented using the
triangle known for applications of artificial cognition [20].
A concept is a link between three notions: the referent, the signifier and the sig-
nified. The referent is what a concept represents, it is here an informative object which
constitutes the class object of a shared information. The informational object is the
basic element for exchanging data through our mediation model. The signifier gathers
all the knowledge allowing to give a meaning to the concept, the taxonomic infor-
mation (synonyms, antonyms), the properties (semantics of values, constraints) and the
relations with the neighborhood. The signified is the conceptual class that represents
the abstraction of an entity.
As SIs can exports mediation model object classes, which provide a local context,
these classes could be defined in relation to their local context, so the mediation model
will describe both the structure and the semantics of the data. In other hand, semantic
reconciliation between the contexts of two SIs allows to import the relevant classes and
integrate them.
3.1 A Cooperation Scenario

This subsection describes an example of interoperability between heterogeneous sys-
tems. An association that wants to provide an information service on the concerts of a
variety artists across Brazil.
S1(nbC, artistN, dateC, freeP, soldP, priceP)

nbC (integer): Identifying Number concert id (local to system)
artistN (string): Artist name
dateC (date): Date of the concert
freeP (integer): Number of free places
soldP (integer): Number of seats sold
priceP (float): Price of a seat (in Brazilian Reals)
S2(id, name, session 1..n, nbP, totP, ticket)
id (integer): Number identifying a concert (local to the system)
name (string): Name of the artist
session 1..n (date): Date of the session
nfP (integer): Number of free places
totP (integer): Total number of places
ticket (float): Price of a seat (in dolars)
S2 has a function of transforming the Brazilian Reals into dolars in the form of
a macro named rtod.
Fig. 1. Schemes of s1 and s2 information systems
The information offered by this association includes for each concert the name of
the artist, the date of the concert, the number of places still free, the price of places and
the room where it takes place.
A user can query this system to obtain an integrated view of the information
distributed on various information sources. To simplify the example, we consider the
existence of two systems s1 and s2 relating to two concert halls and likely to give
information on concerts.
The system s1 is administered by a relational DBMS (a table S1) while s2 uses a

spreadsheet to store the data (a worksheet S2). The organization of the information is
described in Fig. 1.
Sources s1 and s2 pose various problems of heterogeneity. The data models are
different, name conflicts appear between similar entities, value conflicts appear in the
price of places, cognitive conflicts appear in the way of considering concert dates. Each
system begins by exporting the schema of its information in a format specific to our
model.
The local representation of the data is translated into the mediation model in the
form of object classes. A domain specialist constructs the context of the information
source and links classes to that context to define the semantics of each class. The
context can then be recorded at the level of a directory or register.
The association must discover potentially interesting sites and extract the infor-
mation in line with its needs. A domain expert is responsible for defining the context of
the application that becomes the reference context when a system plays the role of
consumer information.
The context specification is compared with all or a subset of the contexts of the sites
registered in the directory. The source contexts are reconciled with the reference
context, the relevant object classes are imported and adapted to the semantics of the
reference context. Imported classes can then be interpreted in the application domain.
Virtual classes provide full access transparency, providing a seamless and inte-
grated interface for shared information. The construction of virtual object classes
consists in identifying the classes of similar or related objects by using the contents of
the reference context and generating matching rules between these classes.
Object classes and integration rules form a static knowledge base for resolving
queries. The dynamic aspect is based on the possibility of incrementally increasing the
content of this knowledge by using contextual reconciliation to import and integrate
new object classes.
3.2 Language and Class Representation

The description of the shared information will use a mediation language that both
describes object classes in an object-oriented model [16, 26] and the semantics of these
classes through a context derived from terminological logic [27].
An object class represents a set of objects that respond to a common structure.
A class is defined by its name and by the list of its methods that characterize the
properties (attributes and operations) of the objects of the class.
The object is represented both by its state (the properties values) and a unique
internal identifier.
Information sources s1 and s2 provide classes cs1 and cs2 for cooperation. The
purely local information such as the identifiers disappear, the attributes of the site
relation s1 and the column headings of the site spreadsheet s2 become the methods of
the classes cs1 and cs2.
The formats of numbers and dates are standardized in the model primitive types.
The representation of the number of numbered sessions of s2 uses the concepts of the
model using a session method set with the session number.
The Brazilian Reals to American dollars conversion function of s2 also becomes a

parameterized method as rtod.
3.3 Context Representation

A context in our model will be a semantic network, that is to say a set of terminological
concepts [23–25] related to each other. To understand a concept is to allow the
interpretation of the classes of objects associated with it. A context is composed of
several semantic elements: the concept that forms the basic semantic unit, the role that
models a binary relation between concepts, specifies relations of generalization/
specialization between the concepts on the one hand and the roles on the other hand,
the taxonomy which defines relations of synonymy and antinomy between the terms of
the terminology of the context, the interpretation which allows an object class to
instantiate a concept.
4 Conclusion and Further Work
In this paper, we have presented an extended scheme mediation integration model for
information systems project proposal. Our approach uses a model that describes the
shared information by taking their semantics into account in the context of usage
contexts. The definition of local contexts location and use of shared and heterogeneous
information turn our work in the direction of further formalizing the notion of semantic
distance and moving towards the implementation of architecture. The next steps of
creating our model will evolve tailoring a methodology based on contextual matching
and flexible architectures which will allow incremental integration of relevant infor-
mation in the context of information systems. Further steps will also be responsible to
formalize the definition of local contexts, contextual matching and semantic distance to
realize the interpretation, finding and use of shared and heterogeneous information. Our
current work goes in the direction of further formalizing the notion of semantic distance
and moving towards the implementation of an architecture.
References
1. Liu, L., Pu, C.: Approach to large-scale interoperable database systems. Technical report
TR95-16, University of Alberta (1995)
2. Floresco, D., Rachid, L., Valdurez, P.: Using heterogeneous equivalences for query rewriting
in multidatabase systems. In: Proceedings of the Third International Conference on
Cooperative Information Systems (1995)
3. Sheth, A.P., Larson, J.A.: Federated database systems for managing distributed, heteroge-
neous, and autonomous databases. ACM Comput. Surv. 22(3), 183 (1990)
4. Wiedershold, G.: Mediators in the architecture of future information systems. IEEE Comput.
Mag. 25(3), 38 (1992)
5. Wiedershold, G.: The basis for mediation. In: Proceedings of the Third International
Conference on Cooperative Information Systems (1995)
6. Cody, W.F., Has, L.M., Nblack, W., Arya, M., Carey, M.J., Fagim, R., Flickner, M., Lee,
D., Petkovick, D., Schwarz, P.M., Thomas, J., Roth, M.T., William, J.H., Wimers, E.L.:
Querying multimedia data from multiple repositories by content: the garlic project. In:
Proceedings of the Third IFIP (1995)
7. Tonasic, A., Rasschid, L., Valdurez, P.: Scaling heterogeneous databases and the design of
disco. In: Proceedings of the IEEE 16th International Conference on Distributed Computing
Systems. IEEE Computer Society Press (1996)
8. Garcia Molina, H., Quas, D., Papakostantinou, Y., Rajarman, A., Sagiv, Y., Ullman, J.D.,
Widon, J.: The TSIMMIS approach to mediation: data models and languages. In:
Proceedings of the Second International Workshop on Next Generation Information
Technologies and Systems (1995)
9. Papakonstatinou, Y., Abitebul, S., Garcia Molina, H.: Object fusion in mediator systems. In:
Proceedings of 24th International Conference on Very Large Data Bases (1996)
10. Kashyap, V., Sheth, A.: Semantics-based information brokering. In: Proceedings of the 3rd
International Conference on Information and Knowledge Management, Gaithersburg. ACM
Press (1994)
11. Bishr, Y.: Semantic aspects interoperable GIS. Ph.D. thesis, Enschede Netherlands (1997)
12. Bressam, S., Goh, C.H., Fiynn, K., Jakobisiack, M., Husein, K., Kon, H.B., Lee, T.,
Maddnick, S.E., Pena, T., Qu, J., Shum, A.W., Siegel, M.: Context interchange mediator
prototype. In: Proceedings of ACM SIGMOD International Conference on Management of
Data (1997)
13. Deker, S., Erdmamn, M., Fensel, D., Studer, R.: Ontobroker: ontology based access to
distributed and semi-structured information. In: Eighth Working Conference on Database
Semantics, DS-8, Kluwer (1999)
14. Grubber, T.: A translation approach to portable ontology specifications. Int. J. Knowl.
Acquisition Knowl. Based Syst. 5(2), 199–220 (1993)
15. Guarino, N.: Semantic matching: formal ontological distinctions for information organiza-
tion, extraction, and integration, summer school on information extraction, Frascati, Italy,
July 1997
16. Kiffer, M., Lausem, G., Wu, J.: Logical foundations of object-oriented and frame-based
languages. J. ACM 42(4), 741 (1995)
17. Mena, E.: OBSERVER: an approach for query processing in global information systems
based on interoperation across pre-existing ontologies. Ph.D. thesis, University of Zaragoza
(1999)
18. Fowler, J., Perry, B., Bargmeyer, B.: Agent-based semantic interoperability in InfoSleuth.
ACM SIGMOD Rec. 28(1), 60 (1999)
19. Nicolle, C., Cullot, N., Yetongnon, K.: A translation process between cooperating
heterogeneous database systems. In: Proceedings of PDCS97, New Orleans (1997)
20. Grumbach, A.: Cognition artificielle. Du réflexe à la réflexion. Addison-Wesley France
(1994)
21. Vitvar, T., Kopecky, J., Viskova, J., Fensel, D.: WSMO-Lite annotations for web services.
In: Bechhofer, S., Hauswirth, M., Hoffmann, J., Koubarakis, M. (eds.) ESWC 2008. LNCS,
vol. 5021, pp. 674–689. Springer, Heidelberg (2008)
22. Farrell, J., Lausen, H. (eds.): Semantic annotations for WSDL and XML schema. W3C
Recommendation (2007)
23. McGuinness, D.L., van Harmelen, F.: OWL web ontology language overview. W3C
Recommendation (2004)
24. Hibner, A., Zielinski, K.: Semantic-based dynamic service composition and adaptation.
IEEE SCW 2007, 213–220 (2007)
25. Gomadam, K., Verma, K., Brewer, D., Sheth, A.P., Miller, J.A.: A tool for semantic
annotation of web services. In: Gil, Y., Motta, E., Benjamins, V.R., Musen, M.A. (eds.)
ISWC 2005, vol. 3729. Springer, Heidelberg (2005)
26. Gmez-Prez, J.M., Erdmann, M., Greaves, M., Corcho, O., Benjamins, R.: A framework and
computer system for knowledge-level acquisition, representation, and reasoning with process
knowledge. Int. J. Hum. Comput. Stud. 68, 641 (2010)
27. Krotzsch, M., Simancik, F., Horrocks, I.: Description logics. IEEE Intell. Syst. 29, 12–29
(2014)
28. Kurniawan, K., Ashari, A.: Service orchestration using enterprise service bus for real-time
government executive dashboard system. In: Proceedings of 2015 International Conference
on Data and Software Engineering, ICODSE 2015 (2016)
29. Parlanti, D., Pettenati, M.C., Bussotti, P., Giuli, D.: Improving information systems
interoperability and flexibility using a semantic integration approach. In: Proceedings - 4th
International Conference on Automated Solutions for Cross Media Content and Multi-
Channel Distribution, Axmedis 2008 (2008)
The 15th International Workshop on
Heterogeneous Wireless Networks
(HWISE-2019)
An Analysis on Recent Mobile Application
Trend in Bangladesh
Md Anik Hasan(&), Nazifa Tasneem, Sumaiya Binte Akther,

Koushiq Das, and Ashik Mostafa Alvi
Department of Electrical and Computer Engineering, North South University,

Plot – 15, Block – B, Bashundhara, Dhaka 1229, Bangladesh
{anik.hasan,nazifa.tasneem173,sumaiya.akther,
koushiq.das,ashik.mostafa}@northsouth.edu
Abstract. Development in the technology sector, making the world closer and
easier day by day. It is very difficult to stay aside without the touch of the
technology. From the invention of computer, the daily life of smart people has
changed. Now we are seeing smart phone is ruling the world. It has replaced the
great invention of computer. Almost every work of a computer and many fea-
tures are now enabled in a small smart phone. The apps inside a smart phone are
one of the main factors that is helping the people to make their life simpler.
Because of the immense usage of smart phone, app developers are increasing
day by day. In this paper, we have tried to find out the recent app using trend in
Bangladesh. We have operated a survey. From there we have figured out the
most demanding app types and also the specification in Bangladesh. Our work
will help the app developers beyond any platform. This will help them to build a
demanding app and be in a good position in the market place.
Keywords: Mobile app Recent app trend Marketplace Smart phone
1 Introduction
The number of mobile phone users in twenty years back and today are so much
amusing that it is pretty much difficult to say how a small device got so much popu-
larity among the all classes of people [1]. In past, it was so costly for a person to buy a
phone. But with time everything has changed. As the cost has went down and the
reachability to phones upgraded, people are willing to buy a phone. The features in a
mobile phone is updating day by day and the technology got a revolutionary change
when it got the taste of first smartphone. How features and technology can be improved
to make the peoples life easy, a smartphone is the best example.
People are getting dependent on tech world to make their life easier and enter-
tainable. In recent times, people want a fast and reliable device for their communication
which would be easy to use as well as the best thing to pass their leisure time. So,
people are changing their choices for communication to smartphones from analog
cellular phones. The main reason of this change is for some eye-catching applications
in the smartphones which are not supported in analog phones. And day by day the

https://doi.org/10.1007/978-3-030-15035-8_18
196 Md A. Hasan et al.
applications are getting more enjoyable and more reliable with the increasing rate of
development of smartphones.
The technology will develop every day and the mobile applications should also be
developed to cope up with the recent running race. In 2014, the number of smartphone
users was 1.57 billion and in 2017, it is 2.32 billion and it is expected that the users will
be 2.87 billion by 2020 [2]. There are many factors behind this increasing number of
users year by year. One of this factor is the mobile applications. People are buying
smartphones to use these applications. So, these applications can be improved or
similar kind of apps can be made.
There are some key factors which are responsible for making a product famous or
popular in the marketplace [3]. The main factors are:
• Knowledge about the targeted product
• Knowledge about the target users’ choice
• Knowledge about the targeted marketplace
Every developer wants to make a good marginable profit from his or her product. It
will be risky to make a product which is not wanted by the user or by a specific user
group. It may cause loss of his or her valuable effort. If a mobile app developer is
willing to build an app then it would be a great loss for the latest up growing
technology.
A mobile application developer needs some knowledge before building a new
app. The first fact to make a new application is the idea because if the app doesn’t get
popular among the users, the effort, time and money everything will be wasted. Maybe
the app is good but the marketplace is not ready for this kind of apps or there can be
another reason that people are wanting something else or there are other similar apps
which are better than this one. And if the developer fails to make his or her app popular,
he or she will lose confidence to work again on this mobile app developing sector.
Our effort is to make the developer’s and freelancers’ life easier. We are focusing to
find out the concurrent technology trend so that a developer can focus only on his or
her app developing work. It will save developer’s valuable time and productivity. By
analyzing the market, we have suggested some most top using apps and their con-
figurations. The developers will have an idea about the app market. They can judge
their idea whether the app they are planning to make will make an impact among the
smartphone users or not. In short, in our work we have achieved:
• Ranked the top usage apps in Bangladesh
• Figured out the category of the top running app
• also figured out which mobile platform is making an impact in profit comparison
• explored out the smartphone configuration which is mostly used by a large number
of people in Bangladesh
• the mobile building giants will also have an idea which configured mobile should be
in the market in near future
The remaining part of this paper is decorated as:
An Analysis on Recent Mobile Application Trend in Bangladesh 197
2 Related Work
The trend of using apps is changing day by day. A certain app is famous for some
certain time. So, the market is fluctuating as the recent technology is jumping to a new
era day by day. So, it is very important to have a good idea about the market to launch a
new product.
There is a research work done on automotive application usage using fuzzy logic
[4]. The focus was mainly on how often the cruise control in cars are used. Summa-
rizing the user’s data into a Cumulative Distribution Function (CDF) and then divided
the speed into four levels, a solution was made using Probability Distribution Function
(PDF) to measure speed preference. And with speed and data the research shows a
result of prediction how often a car owner uses the cruise control in a car.
In a recent research, the categories of mobile applications have been studied [5]. In
that work, four categories are mentioned for various mobile apps. These are - a)
notifications, b) location tracking, c) navigation, d) real-time mobile job scheduling.
Similarly, Habib, Ateeq and Bajwa conducted a survey for examining the basic
trends of the usage of mobile applications and different methods for their allocation on
the mobile phones. To get their desired result, they asked four simple questions through
their survey. The outcome of the survey was successful to measure the popularity of
various internet applications and the distribution methods used for the mobile appli-
cations [6].
Moreover, in another research paper, the authors made an app named ‘Analysis and
Prediction of Application Usage’ (APAU) where their main target was to make
application switching in an android phone smooth, fast and efficient by helping the user
conserve time and effort, which could otherwise be wasted on work other than the
actual usage [7]. They resolved to build a recommendation system that will point out
applications based on the user’s usage pattern. Users can be helped by APAU because
it can accelerate application usage by providing shortcuts to applications that the user
would want to use next.
In the generation we are living, there are tremendous number of apps we can choose
from. People from different sectors can easily access the apps according to their need.
In a work, the researchers had conducted a survey to get the users requirements of an
online course management system [15]. Feng, Worrachananun and Lai proposed a
research based on the people who use apps for their learning purpose. Their study has
analyzed learner’s learning styles and better understands their perceptions on the use of
a smartphone application based on the unified theory of acceptance and use of tech-
nology [8]. Interestingly, they discovered that most of the learners prefer visual
learning style, diagrams and graphs which can help them to gather information and
memorize things.
Similarly, in another research, Diliberto-Macaluso and Hughes brought in a
resembling idea with a little modification. The goal of their experiment was to deter-
mine whether using mobile apps in class had any educational advantage. The results
demonstrated one way in which mobile devices, in general, and mobile apps, specif-
ically, can be effectively integrated in an introduction to psychology class to promote
student learning. Their findings suggested that using mobile apps in class could serve
as a useful pedagogical aid to enhance student learning [9].
Apart from these, a vision paper was done for making an impact on developers
from both small and large organizations to develop mobile apps. They mainly focused
on some software development phases (i.e., requirements, security, development,
testing etc.) and discussed about the challenges, future directions and risks of each of
the phases. Their expectance was that the research paper could assist newcomers to
quickly gain a background in the area of mobile apps. Moreover, they hoped that the
argument of the vision for the area will inspire and guide future work and build a
community of researchers with common goals regarding software engineering chal-
lenges for mobile apps [10].
In addition, in a paper, a group of researchers proposed that they had created a
mobile app for the developers to reduce failure of apps as number of failing apps is
growing enormously. They denoted four reasons for failing apps (role of the developer,
user, technical details and marketing) and suggestions to improve them. The initial
motive of this research was to show that the methodological aspect of software engi-
neering can be successfully applied to mobile applications, with the help of software
development life cycle [11]. They used UML, M-UML and other mobile application
development technologies for this purpose.
One more app named “Gate 2015” [12], is available having modules for syllabus,
practice questions, previous question papers, reference of subjects, guidelines etc.
A similar app called “Net Prep+” [13], is available having modules (i.e., Timed test,
Random test, Chapter Wise test etc.). Likewise, another app named “iquiz” was
developed by Kathuria and Gupta that could help students in preparation of competitive
exams like- UGC-NET, GATE etc. Their vision paper also highlighted several chal-
lenges faced by developers in Android App Development [14].
3 Methodology
In this section, we have presented our own approach to find out the specifications for
top mobile applications. For that reason, we have to conduct a survey to collect the
data.
3.1 The Survey: Dataset

We had set six questions to get our data. The targeted people of our survey was within
the capital of Bangladesh: Dhaka city. Most of the data provider were university
students. In Bangladesh, most of the smart phone users are young people. The survey
questions are related to:
• Gender (Male/Female)
• Age Range (10–18/19–40/40+ years)
• Smart Phone Name with Model
• Smart Phone’s OS
• Internal Storage (8/16/32/64/128/256 GB)
• Smart Phone’s Processor
• List of Apps Installed in the Phone
There were 943 people who had helped us by providing their smart phone’s
information. Among them, 688 people were male and 255 were female. And from our
survey we have found out that most of the smart phone users (827 people) are from 19
to 40 years old. Figure 1 is the overview of our dataset.
Fig. 1. Dataset overview.
3.2 Data Processing

The raw data that we have gathered by the survey had some outliers. Some of the data
fields were missing for some individual survey providers. We have filled those data by
putting the average value of that corresponding field.
We also have combined three data fields together and created a new field named
Specification. The processor, ram and internal storage are combined together and called
as specification. And from the data we have collected, we have found that there are 43
specifications. But we have taken 40 specifications for our work. Three specifications
are not counted because those are outdated and their supports from their phone com-
pany is closed and also there are few apps running within those phones.
Then we have separated all the apps and counted total number of times each of the
apps are present in a particular specification. We have discarded the apps which are
installed less than ten specifications. And finally, we have found fifteen apps which are
frequently used and installed within those forty specifications. We have also mapped
down the specifications with numeric value starting from one to forty. Figure 2 rep-
resents the specifications mapping.
Fig. 2. Specification mapping.
Figure 3 shows the final dataset before our work begins.
Fig. 3. Filtered dataset.
4 Our Findings
Figure 4(a) shows Facebook app is mostly used in specification number 26 which is
processor: Quad Core, internal storage: 16 GB and RAM: 1.5 GB. It has appeared 120
times in specification number 26’s phone.
(a) (b) (c)
Fig. 4. Top 1–3 apps (Facebook, Instagram, Messenger).

Figure 4(b) shows Instagram app is mostly used in specification number 5 which is
processor: Octa Core, internal storage: 64 GB and RAM: 4 GB and 12 which is
processor: Octa Core, internal storage: 32 GB and RAM: 4 GB. It has appeared 30
times in specification number 5 & 12’s phone.
Figure 4(c) shows Facebook Messenger app is mostly used in specification number
10 which is processor: Octa Core, internal storage: 16 GB and RAM: 2 GB, 5: pro-
cessor: Octa Core, internal storage: 64 GB and RAM: 4 GB and 12: processor: Octa
Core, internal storage: 32 GB and RAM: 4 GB. It has appeared 36, 30 and 30 times in
specification number 10, 5 and 12’s Phone respectively.
(a) (b) (c)
Fig. 5. Top 4–6 apps (Uber, Imo, Viber).
Figure 5(a) shows Uber app is mostly used in specification number 5: processor:
Octa Core, internal storage: 64 GB and RAM: 4 GB, 12: processor: Octa Core, internal
storage: 32 GB and RAM: 4 GB and 19: processor: Quad Core, internal storage:
16 GB and RAM: 2 GB. It has appeared 42, 30 & 36 times in specification number 5,
12 & 19’s phone respectively.
Figure 5(b) shows Imo app is mostly used in specification number 10: processor:
Octa Core, internal storage: 16 GB and RAM: 2 GB, 19: processor: Quad Core,
internal storage: 16 GB and RAM: 2 GB and 32: processor: Quad Core, internal
storage: 8 GB and RAM: 2 GB. It has appeared 30, 24 & 18 times in specification
number 10, 19 & 32’s phone respectively.
Figure 5(c) shows Viber app is mostly used in specification number 5: processor:
Octa Core, internal storage: 64 GB and RAM: 4 GB and 18: processor: Octa Core,
internal storage: 64 GB and RAM: 4 GB and 19: processor: Quad Core, internal
storage: 16 GB and RAM: 2 GB. It has appeared 18 times in specification number 5
and 19’s phone.
(a) (b) (c)
Fig. 6. Top 7–9 apps (Snapchat, Color Switch, SHAREit).
Figure 6(a) shows Snapchat app is mostly used in specification number 2: pro-
cessor: Octa Core, internal storage: 32 GB and RAM: 3 GB. It has appeared 54 times
in specification number 2’s phone.
Figure 6(b) shows Color Switch app is mostly used in specification number 2:
processor: Octa Core, internal storage: 32 GB and RAM: 3 GB, 20: processor Octa
Core, internal storage: 64 and RAM: 1.5 GB. It has appeared 6 times in specification
number 2 and 20’s phone.
Figure 6(c) shows SHAREit app is mostly used in specification number 2: pro-
cessor: Octa Core, internal storage: 32 GB and RAM: 3 GB and 10 which is processor:
Octa Core, internal storage: 16 GB and RAM: 2 GB. It has appeared 36 and 18 times
in specification number 2 & 10’s phone.
(a) (b) (c)
Fig. 7. Top 10–12 apps (Pathao, Snapseed, Chrome).
Figure 7(a) shows Pathao app is mostly used in specification number 9: processor:
Dual Core, internal storage: 64 GB and RAM: 2 GB and 12 which is processor: Octa
Core, internal storage: 32 GB and RAM: 4 GB. It has appeared 18 times in specifi-
cation number 9 & 12’s phone.
Figure 7(b) shows Snapseed app is mostly used in specification number 2: pro-
cessor: Octa Core, internal storage: 32 GB and RAM: 3 GB. It has appeared 42 times
in specification number 2’s phone.
Figure 7(c) shows Chrome app is mostly used in specification number 2: processor:
Octa Core, internal storage: 32 GB and RAM: 3 GB and 9 which is processor: Dual
Core, internal storage: 64 GB and RAM: 2 GB. It has appeared 18 times in specifi-
cation number 2 and 9’s phone.
5 Conclusion
The given facts and figures should be enough to exhibit the applications generally
being used on smart phones lately. Top-notch industries or high-tech software com-
panies are generally looking for such statistics in order to perceive the types of apps
and specifications on demand. This can save time and unnecessary effort on making
excess apps that are redundant.
App developers and freelancers of our country can also be highly benefited from all
this information given, as it is very important for them to invest their resources in the
right places. Moreover, this methodology can be implemented in other countries. If a
created app does not thrive on the market, such small town businesses or home-based
workshops can suffer extreme loss.
References
1. Eason, G., Noble, B., Sneddon, I.N.: On certain integrals of Lipschitz-Hankel type involving
products of Bessel functions. Philos. Trans. R. Soc. London A247, 529–551 (1955)
2. Clerk Maxwell, J.: A Treatise on Electricity and Magnetism, vol. 2, 3rd edn., pp. 68–73.
Clarendon, Oxford (1892)
3. Jacobs, I.S., Bean, C.P.: Fine particles, thin films and exchange anisotropy. In: Rado, G.T.,
Suhl, H. (eds.) Magnetism, vol. III, pp. 271–350. Academic, New York (1963)
4. Qiu, S., McGee, R.A., Murphey, Y.L.: Adaptive fuzzy prediction for automotive
applications usage. In: 2015 IEEE 14th International Conference on Machine Learning
and Applications (ICMLA) (2015). https://doi.org/10.1109/icmla.2015.138
5. Zhao, F., Shin, J., Reich, J.: Information-driven dynamic sensor collaboration. IEEE Sig.
Process. Mag. 19(2), 61–72 (2002). https://doi.org/10.1109/79.985685
6. Habib, H., Ateeq, M., Bajwa, I.S.: A study on trends of mobile application usage and their
distribution methods in Pakistan. In: Eighth International Conference on Digital Information
Management (ICDIM 2013) (2013). https://doi.org/10.1109/icdim.2013.6693976
7. Acharya, S., Shenoy, A., Lewis, M., Desai, N.: Analysis and prediction of application usage
in android phones. In: 2016 2nd International Conference on Advances in Electrical,
Electronics, Information, Communication and Bio-informatics (AEEICB) (2016). https://doi.
org/10.1109/aeeicb.2016.7538346
8. Feng, Y., Worrachananun, M., Lai, I.K.: Students preferences and intention on using
smartphone education applications. In: 2015 International Symposium on Educational
Technology (ISET) (2015). https://doi.org/10.1109/iset.2015.30
9. Diliberto-Macaluso, K., Hughes, A.: The use of mobile apps to enhance student learning in
introduction to psychology. Teach. Psychol. 43(1), 48–52 (2015). https://doi.org/10.1177/
0098628315620880
10. Nagappan, M., Shihab, E.: Future trends in software engineering research for mobile apps.
In: 2016 IEEE 23rd International Conference on Software Analysis, Evolution, and
Reengineering (SANER) (2016). https://doi.org/10.1109/saner.2016.88
11. Inukollu, V.N., Keshamon, D.D., Kang, T., Inukollu, M.: Factors influencing quality of
mobile apps: role of mobile app development life cycle. Int. J. Softw. Eng. Appl. 5(5), 15–34
(2014). https://doi.org/10.5121/ijsea.2014.5502
12. Surachat, K., Kajkamhaeng, S., Damkliang, K., Tiprat, W., Wacharanimit, T.: First aid
application on mobile device. Int. J. Comput. Inf. Eng. 7(5), 622–627 (2013). https://waset.
org/publications/13742/first-aid-application-on-mobile-device. Accessed 20 Apr 2018
13. Joorabchi, M.E., Mesbah, A., Kruchten, P.: Real challenges in mobile app development. In:
2013 ACM/IEEE International Symposium on Empirical Software Engineering and
Measurement (2013). https://doi.org/10.1109/esem.2013.9
14. kathuria, A., Gupta, A.: Challenges in Android application development: a case study. Int.
J. Comput. Sci. Mobile Comput. 4(5), 294–299 (2015). https://www.ijcsmc.com/docs/
papers/May2015/V4I5201557.pdf. Accessed Apr 20
15. Alvi, A.M., Shaon, M.F., Das, P.R., Mustafa, M., Bari, M.R.: Automated course
management system. In: 2017 12th International Conference for Internet Technology and
Secured Transactions (ICITST) (2017). https://doi.org/10.23919/icitst.2017.8356369
Performance Evaluation of Routing
Protocols in DTNs Considering Different
Mobility Models
Evjola Spaho1(B) , Klodian Dhoska2 , Kevin Bylykbashi3 , Leonard Barolli4 ,

Vladi Kolici1 , and Makoto Takizawa5
1
Department of Electronics and Telecommunication,
Faculty of Information Technology, Polytechnic University of Tirana,
Mother Teresa Square, No. 4, Tirana, Albania
evjolaspaho@hotmail.com, vkolici@fti.edu.al
2
Department of Production-Management, Faculty of Mechanical Engineering,
Polytechnic University of Tirana, Mother Teresa Square, No. 4, Tirana, Albania
kdhoska@upt.al
3
Graduate School of Engineering, Fukuoka Institute of Technology (FIT),
3-30-1 Wajiro-Higashi, Higashi-Ku, Fukuoka 811-0295, Japan
kevini 95@hotmail.com
4
Department of Information and Communication Engineering,
Fukuoka Institute of Technology (FIT), 3-30-1 Wajiro-Higashi,
Higashi-Ku, Fukuoka 811-0295, Japan
barolli@fit.ac.jp
5
Department of Advanced Sciences, Hosei University, 3-7-2, Kajino-cho,
Koganei-shi, Tokyo 184-8584, Japan
Abstract. In this paper we evaluate the performance of Epidemic,

Spray and Wait routing protocols and their versions with congestion
control and Epidemic with TCP in Delay Tolerant Networks. For evalua-
tion we used three different mobility models: Random Waypoint (RWP),
Steady State Random Waypoint (SSRWP) and a Tirana city map based
movement. We used delivery ratio, hop count, average delay and average
buffer occupancy metrics to evaluate the network performance. The net-
work performs better for SSRWP compared to RWP and realistic Tirana
map-based scenarios. The Epidemic with TCP has a high average delay
because of ack packets.
1 Introduction
Nowdays the number of mobile devices is highly increased. Users can access
information and communicate with each other anywhere and anytime. In high
mobility networks and long distances, end to end connectivity is difficult to
maintain with current technology. In order to deal with this problem, Delay
Tolerant Networks (DTNs) can be used. In DTNs, mobile nodes can send and
receive data, carry data as relays and forward data in opportunistic way upon
https://doi.org/10.1007/978-3-030-15035-8_19
206 E. Spaho et al.
contacts. In order to handle disconnections and long delays, DTNs use store-
carry-and-forward approach.
Instead of working end-to-end, in VDTNs, a message-oriented overlay
layer called Bundle layer employs a store-carry-and-forward message switching
paradigm that moves messages from node to node, along a path that eventually
reaches the destination.
In DTNs multiple copy routing protocols are mostly used. Multiple copies
can lead to congestion in networks that are typically characterized by scarce
resources, so a message deletion policy must be used for multi copy routing
schemes. These schemes can use the acknowledgement method to remove the
useless bundles from the network, preventing the nodes from the buffer overflow
problem and avoid transfer of useless message replicas thus relaxing the resources
of the nodes.
Combining routing with a congestion control scheme would manage better
nodes buffer. In this work, we use the congestion control mechanism for DTNs
proposed in [1] and [2]. This mechanism locally advertises the node’s buffer
occupancy to adjacent nodes in order to avoid sending messages to nodes whose
buffers are nearly full. Congestion control mechanism can maximize resource
utilization when resources are available.
In this paper, we investigate and compare the performance of Epidemic,
Spray and Wait and their versions with congestion control and Epidemic with
TCP. Random Waypoint (RWP), Steady State Random Waypoint (SSRWP) and
Tirana map-based mobility models are used. Delivery ratio, hop count, average
delay and average buffer occupancy metrics are used to evaluate the performance.
Simulations are conducted with Network Simulation 3 (ns-3) [3].
The remainder of this paper is as follows. Section 2 introduces DTNs and
routing protocols. The simulation system design and simulation scenarios are
presented in Sect. 3. In Sect. 4 are shown the simulation results. Finally, the
conclusions and future work are presented in Sect. 5.
2 DTNs and Routing Protocols

2.1 DTN Overview
DTN is the “challenged computer network” approach that is originally designed

from the Interplanetary Internet, and the data transmission is based upon the
store-carry-and-forward protocol for the sake of carrying data packets under a
poor network environment such as space [4].
In large networks where a continuous path between the source and destination
does not exist [4,5], DTNs can be used. In DTNs, data can be transmitted by
storing them at nodes and forwarding them later when a link is established. This
technique is called message switching. By using opportunistic contacts, data will
be relayed to the destination.
Different copies of the same bundle can be routed independently to desti-
nation, thus improving the delivery probability and reducing the delivery delay.
Performance Evaluation of Routing Protocols in DTNs 207
However, when a big number of copies are created in the network, such approach
increases the contention for network resources as bandwidth and storage.
In [6], authors have studied this model and found that it can provide sub-
stantial capacity at little cost, and that the use of a DTN model often doubles
that capacity compared with a traditional end-to-end model. The main assump-
tion in the Internet that DTNs seek to relax is that an end-to-end path between
a source and a destination exists for the entire duration of a communication
session. When this is not the case, the normal Internet protocols fails. DTNs get
around the lack of end-to-end connectivity with an architecture that is based on
message switching. It is also intended to tolerate links with low reliability and
large delays. The architecture is specified in RFC 4838 [7].
Bundle protocol has been designed as an implementation of the DTN archi-
tecture. A bundle is a basic data unit of the DTN bundle protocol. Each bundle
comprises a sequence of two or more blocks of protocol data, which serve for
various purposes. In poor conditions, bundle protocol works on the application
layer of some number of constituent Internet, forming a store-and-forward over-
lay network to provide its services. The bundle protocol is specified in RFC 5050.
It is responsible for accepting messages from the application and sending them
as one or more bundles via store-carry-and-forward operations to the destination
DTN node. The bundle protocol runs above the TCP/IP level.
2.2 Routing in DTNs
In order to handle disconnections and long delays in sparse network scenarios,

DTN uses store-carry-and-forward approach. A network node stores a bundle and
waits for a future opportunistic connection. When the connection is established,
the bundle is forwarded to an intermediate node, according to a hop-by-hop
forwarding/routing scheme. This process is repeated and the bundle will be
relayed hop-by-hop until reaching the destination node. A number of different
routing protocols have been proposed for DTNs.
Depending on the number of copies they create, DTN routing protocols
are classified in single-copy and multiple-copy. As the name suggests in single-
copy routing protocols, only one copy of each message exists in the network. In
multiple-copy routing protocols there are two subgroups: n-copy routing proto-
cols and unlimited copy routing protocols. N-copy protocols limits the number of
message copies created to a configurable maximum and distributes these copies
to contacts until the number of copies is exhausted. Unlimited copy routing
protocols are variants of flooding.
In [8–23], authors deal with routing in DTNs. In this work, we will use two
widely applicable DTN routing protocols Epidemic [20] and Spray and Wait [21].
Epidemic Routing Protocol: Epidemic [20] is a protocol that is basically a

flooding mechanism. Each message spreads like a disease in a population without
priority and without limit. When two nodes encounter each other they exchange
a list of message IDs and compare those IDs to decide which message is not
208 E. Spaho et al.
already in storage in the other node. The next phase is a check of available
buffer storage space, with the message being forwarded if the other node has
space in its buffer storage. The main goals of this protocol are: maximize the
delivery ratio, minimize the latency and minimize the total resources consumed
in message delivery. It is especially useful when there is lack of information
regarding network topology and nodes mobility patterns.
Spray and Wait Routing Protocol: Spray and Wait [21], is a routing proto-
col that attempts to gain the delivery ratio benefits of replication-based routing
as well as the low resource utilization benefits of forwarding-based routing. The
Spray and Wait protocol is composed of two phases: the spray phase and the wait
phase. When a new message is created in the system, a number L is attached
to that message indicating the maximum allowable copies of the message in the
network. During the spray phase, the source of the message is responsible for
“spraying”, or delivery, one copy to L distinct “relays”. When a relay receives
the copy, it enters the wait phase, where the relay simply holds that particular
message until the destination is encountered directly.
3 Simulation Setup
3.1 Simulation System
The simulations are conducted on ns-3 (ver-3.26) simulation system. ns-3 simu-
lator is a discrete-event network simulator for Internet systems [3]. ns-3 provides
models of how packet data networks work and perform, and provides a simu-
lation engine for users to conduct simulation experiments. Scripting in ns-3 is
done in C++ or Python.
The ns-3 simulator has models for all network elements that comprise a
computer network. For example, network devices represent the physical device
that connects a node to the communication channel. This might be a simple
Ethernet network interface card, or a wireless IEEE 802.11 device.
The ns-3 support standardized output formats for trace data, such as the
pcap format used by network packet analyzing tools. Several external animators
and data analysis and visualization tools can be used with ns-3.
3.2 Simulation Scenarios and Parameters

We created three different simulation scenarios. In the first scenario is used
RWP mobility model. In the second scenario is used SSRWP and in the third
one Tirana map-based mobility model. In all scenarios the number of nodes
is considered 100 nodes and the nodes move in an area size 2.5 km × 2.5 km.
The simulation time is 1800 s. Destination nodes generate return receipts to the
source node and they are used as antipackets. Antipackets delete the bundle
copies in the buffer in order to make free space. Antipackets have a lifetime of
1000 s, a size of 10 Bytes and a 26 bytes header.
Table 1. Simulation parameters and their values.
Parameters Values
Number of nodes 100
Simulation time 1800 s
Map size 2.5 km × 2.5 km
Movement Model RWP, SSRWP, Map-based
Antipackets lifetime 10000 s
Antipackets size 10 Bytes
Antipackets header 26 bytes
Congestion threshold 0.7
Bundle lifetime 750 s
Bundle retransmission timeout 1000 s
Number of bundle retransmission 3
Max number of bundle copies for Spray and Wait 16
Hello message interval 100 ms
For all scenarios we evaluate the performance of Epidemic and Spray and
Wait protocols and their versions with congestion control (congestion threshold
0.7) and also use Epidemic with TCP. A bundle can be generated only if the
node has sufficient amount of free buffer space. DTN nodes advertise their buffer
content to each other every 100 ms by sending Hello messages. Moreover, if DTN
congestion control is enabled, a bundle can be forwarded only to such interme-
diate nodes that have enough free buffer space. Simulation parameters and their
values are shown in Table 1.
In following, we will briefly describe different mobility models considered in
this work. The RWP model is a commonly used mobility model for simulations
of wireless communication networks. Basically, in RWP every node picks up a
random destination and a random velocity at certain points called waypoints.
This model has been extended in a number of ways in order to take into account
more realistic movements.
SSRWP model [24,25] is based on RWP model for case when speed, pause
and position are uniformly distributed random variables. The difference is that
the initial values of these parameters are not from uniform distribution but from
stationary distribution of RWM model.
Tirana map-based mobility model is a realistic scenario with vehicles moving
in Tirana city created by using Open Street Map (OSM) [26] and Simulation of
Urban MObility (SUMO) [27] (see Figs. 1 and 2).
4 Simulation Results
The simulation results of delivery ratio for RWP, SSRWP and Tirana map-
based mobility model are shown in Fig. 3. Delivery ratio is calculated as the
210 E. Spaho et al.
Fig. 1. Tirana city map from OSM.
Fig. 2. Tirana city map generated from SUMO.
ratio of number of delivered messages to that of created messages. Under the

same simulation conditions SSRWP has higher delivery ratio compared with
two other mobility models. This is related with the probabilistic distribution
with steady state for the initial node position and speed, because SSRWP uses
different distribution from RWP (RWP uses uniform distribution). In SSRWP
the initial time for setting them is not needed before the simulations. When
SSRWP is used the network created will be more stable, more contacts between
nodes can occur and more bundles can be delivered. The decrease of delivery
ratio for Epidemic with TCP shows that SSRWP is better for one directional
communications rather than two directional communication. Congestion control
improves the delivery ratio for both protocols because bundles are not generated
if the buffers are full.
In Fig. 4 are shown the results of the average delay. Average delay is calcu-
lated as the average time elapsed from the creation of the messages at source to
their successful delivery to the destination. Decrease of average delay for SSRWP
1
Binary Spray and Wait
Binary Spray and Wait+CC and Tc=0.7
Epidemic
Epidemic+CC and Tc=0.7
Epidemic+TCP
0.8
Delivery ratio
0.6
0.4
0.2
0
RWP SSRWP Tirana Map-based
Mobility Model
Fig. 3. Simulation results of delivery ratio.
600
Epidemic
Epidemic+TCP
500
400
Delay (s)
300
200
100
0
Mobility Model
Fig. 4. Simulation results of delay.
5
Epidemic
Epidemic+TCP
4
Hop number
0
Mobility Model
Fig. 5. Simulation results of hop count.

212 E. Spaho et al.
3x106 Binary Spray and Wait

Epidemic
Epidemic+TCP
6
2.5x10
Buffer occupancy
6
2x10
1.5x106
1x106
500000
0
Mobility Model
Fig. 6. Simulation results of buffer occupancy.
shows that SSRWP performs better than RWP. For Spray and Wait and Epi-
demic with and without congestion control the delays are almost the same. For
Epidemic with TCP, the delay is higher because of acks.
The results for hop count are shown in Fig. 5. Hop count is the average
number of hops counts between the source and destination. Hop count is almost
the same for both scenarios because the number of nodes is the same. Epidemic
protocol will have higher hop count because it uses more network resources
(intermediate nodes) compared with Spray and Wait.
In Fig. 6 are shown the results for average buffer occupancy. Average buffer
occupancy shows the average amount of bytes stored in the buffer nodes. For
SSRWP, we can see a decrease of buffer occupancy for all protocols. If we have
a node “a” that have met before node “b” than it has more chances to meet it
again compared with RWP. In the encounters after the first time, nodes exchange
also antipackets that delete duplicated bundles from the buffers. The importance
of congestion control is that it controls the buffers from being full. The usage of
congestion control and the good management of buffers leads to better delivery
ratio. For Epidemic with TCP, the buffer occupancy is higher than Spray and
Wait because more packets are created.
5 Conclusions
In this paper, we evaluated and compared the performance of Epidemic, Spray

and Wait, their versions with congestion control and Epidemic with TCP when
SSRWP, RWP and map based mobility models are used. From the simulation
results, we conclude as follows.
• The protocols performance using SSRWP and RWP is higher compared with
the realistic scenarios because in the realistic scenario the nodes movement is
more complex and the opportunistic contacts are more rare.
• When SSRWP is used the network will be more stable, more contacts between
nodes can occur and more bundles can be delivered compared with RWP.
• Congestion control improves the delivery ratio and buffer management for
both protocols. It uses a mechanism to stop bundle generation and save buffers
from being full.
• The usage of Epidemic with TCP leads to high average delay in all scenarios
because of ack packets.
In the future, we would like to make extensive simulations to evaluate the

performance of different routing protocols, mobility models and congestion con-
trol mechanisms considering new scenarios and parameters.
References
1. Lakkakorpi, J., Ginzboorg, P.: ns-3 module for routing and congestion control
studies in mobile opportunistic DTNs. In: Proceedings of Performance Evaluation
of Computer and Telecommunication Systems (SPECTS), pp. 1–5 (2013)
2. Lakkakorpi, J., Pitkanen, M., Ott, J.: Using buffer space advertisements to avoid
congestion in mobile opportunistic DTNs. In: Proceedings of WWIC, Barcelona,
Spain, pp. 1–12, June 2011
3. The ns-3 Network Simulator. http://www.nsnam.org
4. Fall, K.: A delay-tolerant network architecture for challenged Internets. In: Pro-
ceedings of the International Conference on Applications, Technologies, Architec-
tures, and Protocols for Computer Communications, ser. SIGCOMM 2003, pp.
27–34 (2003)
5. Delay- and disruption-tolerant networks (DTNs) tutorial, NASA/JPL’s Inter-
planetary Internet (IPN) Project (2012). http://www.warthman.com/images/
DTNTutorialv2.0.pdf
6. Laoutaris, N., Smaragdakis, G., Rodriguez, P., Sundaram, R.: Delay tolerant bulk
data transfers on the Internet. In: Proceedings of the 11th International Joint
Conference on Measurement and Modeling of Computer Systems (SIGMETRICS
2009), pp. 229–238 (2009)
7. Cerf, V., Burleigh, S., Hooke, A., Torgerson, L., Durst, R., Scott, K., Fall, K., Weiss,
H.: Delay-tolerant networking architecture. IETF RFC 4838 (Informational), April
2007
8. Massri, K., Vernata, A., Vitaletti, A.: Routing protocols for delay tolerant net-
works: a quantitative evaluation. In: Proceedings of ACM workshop PM2HW2N
2012, pp. 107–114 (2012)
9. Massri, K., Vitaletti, A., Vernata, A., Chatzigiannakis, I.: Routing protocols for
delay tolerant networks: a reference architecture and a thorough quantitative
evaluation. J. Sens. Actuator Networks, 1–28 (2016). https://doi.org/10.3390/
jsan5020006
10. Demmer, M., Fall, K.: DTLSR: delay tolerant routing for developing regions. In:
Proceedings of the 2007 ACM Workshop on Networked Systems for Developing
Regions, 6 p. (2007)
11. Ilham, A.A., Niswar, M., Agussalim: Evaluated and optimized of routing model on
Delay Tolerant Network (DTN) for data transmission to remote area. In: Proceed-
ings of FORTEI, Indonesia University Jakarta, pp. 24–28 (2012)
214 E. Spaho et al.
12. Uchida, N., Ishida, T., Shibata, Y.: Delay tolerant networks-based vehicle-to-
vehicle wireless networks for road surveillance systems in local areas. Int. J. Space-
Based Situated Comput. 6(1), 12–20 (2016)
13. Bylykbashi, K., Spaho, E., Barolli, L., Xhafa, F.: Routing in a many-to-one com-
munication scenario in a realistic VDTN. J. High Speed Networks 24(2), 107–118
(2018)
14. Bylykbashi, K., Spaho, E., Barolli, L., Xhafa, F.: Impact of node density and TTL
in vehicular delay tolerant networks: performance comparison of different routing
protocols. Int. J. Grid Util. Comput. 7(3), 136–144 (2017)
15. Jain, S., Fall, K., Patra, R.: Routing in a delay tolerant network. In: Proceedings of
ACM SIGCOMM 2004 Conference on Applications, Technologies, Architectures,
and Protocols for Computer Communication, Portland, Oregon, USA, 30 August–3
September 2004, pp. 145–158 (2004)
16. Zhang, Z.: Routing in intermittently connected mobile ad hoc networks and delay.
IEEE Commun. Surv. Tutorials 8(1), 24–37 (2006)
17. Soares, V.N.G.J., Rodrigues, J.J.P.C., Farahmand, F.: GeoSpray: a geographic
routing protocol for vehicular delay-tolerant networks. Inf. Fusion 15(1), 102–113
(2014)
18. Burgess, J., Gallagher, B., Jensen, D., Levine, B.N.: Maxprop: routing for vehicle-
based disruption-tolerant networks. In: Proceedings of the IEEE Infocom, April
2006
19. Lindgren, A., Doria, A., Davies, E., Grasic, S.: Probabilistic routing protocol for
intermittently connected networks. draft-irtf-dtnrg-prophet-09. http://tools.ietf.
org/html/draft-irtf-dtnrg-prophet-09
20. Vahdat, A., Becker, D.: Epidemic routing for partially connected ad hoc networks.
Technical report CS-200006, Duke University, April 2000
21. Spyropoulos, T., Psounis, K., Raghavendra, C.S.: Spray and Wait: an efficient
routing scheme for intermittently connected mobile networks. In: Proceedings of
ACM SIGCOMM 2005 - Workshop on Delay Tolerant Networking and Related
Networks (WDTN-05), Philadelphia, PA, USA, pp. 252–259 (2005)
22. Bylykbashi, K., Spaho, E., Barolli, L., Takizawa, M.: Comparison of spray and
wait and epidemic protocols in different DTN scenarios. In: Proceedings of the
12th International Conference on Broad-Band Wireless Computing, Communica-
tion and Applications (BWCCA-2017), pp. 218–229 (2017)
23. Spaho, E., Bylykbashi, K., Barolli, L., Takizawa, L.: Routing in a DTN: perfor-
mance evaluation for random waypoint and steady state random waypoint using
NS3 simulator. In: Proceedings of the 12th International Conference on P2P, Par-
allel, Grid, Cloud and Internet Computing (3PGCIC-2017), pp. 133–141 (2017)
24. Navidi, W., Camp, T.: Stationary distributions for the random waypoint mobility
model. IEEE Trans. Mob. Comput. 3(1), 99–108 (2004)
25. Navidi, W., Camp, T., Bauer, N.: Improving the accuracy of random waypoint
simulations through steady-state initialization. In: Proceedings of the 15th Inter-
national Conference on Modeling and Simulation (MS 2004), pp. 319–326, March
2004
26. Open street map. http://www.openstreetmap.org/
27. Simulation of urban mobility. http://sumo.sourceforge.net/
A Distance-Based
Advertisement-Delivery Method
for Vehicular DTN
Shogo Nakasaki1 , Yu Yoshino1 , Makoto Ikeda2(B) , and Leonard Barolli2

1
Graduate School of Engineering, Fukuoka Institute of Technology,
3-30-1 Wajiro-higashi, Higashi-ku, Fukuoka 811-0295, Japan
tshogonakasakit@gmail.com, mgm17107@bene.fit.ac.jp
2
Fukuoka Institute of Technology,
3-30-1 Wajiro-higashi, Higashi-ku, Fukuoka 811-0295, Japan
makoto.ikd@acm.org, barolli@fit.ac.jp
Abstract. In this paper, we propose a distance-based message relay-

ing method to limit the replicated bundle messages in Vehicular Delay/
Disruption/Disconnection Tolerant Networking (DTN). From the simu-
lation results, we found that our proposed recovery method limits the
duplicated bundle messages considering specified area by applying the
distance-based advertisement-delivery method.
Keywords: Message relaying method · Vehicular DTN ·

Distance-based advertisement-delivery
1 Introduction
Recently, many vehicle companies have been developing vehicles together with
ICT companies to provide the added value of services. Traditionally, vehicles
have been used for people moving and baggage transfer, but until now vehicles
are going to be connected to the network and becoming mobile terminals [4,7,12].
Vehicles will connect to various wireless networks that will become relay nodes
and will have various applications [9,13,21,23,27]. The vehicles will be equipped
with On-Board-Unit (OBU) by plug-in. However, there are many problems that
used to be solved.
Delay/Disruption/Disconnection Tolerant Networking (DTN) is a message
relaying method in Vehicular to Vehicular (V2V) communication [11,17,19,20,
22]. In Vehicular DTN, consumption of network resources and storage usage of
each node becomes a critical problem due to the DTN nodes duplicate messages
to other nodes.
In our previous work [10], we have proposed an Enhanced Message Suppres-
sion Controller (EMSC) for Vehicular-DTN. We have shown that the EMSC
can reduce the consumption of network resources. But, EMSC needs some addi-
tional control messages to reduce the replicated bundle messages. In addition,
https://doi.org/10.1007/978-3-030-15035-8_20
216 S. Nakasaki et al.
we proposed a threshold-based adaptive method [26] for message suppression in

Vehicular DTNs. The proposed adaptive method decreased the replicated bun-
dle messages compared with conventional method, but the distance information
was not considered.
In this paper, we propose a distance-based message relaying method to limit
the replicated bundle message. We evaluate the proposed method by simulations
and consider position distribution and delay as evaluation metrics.
The structure of the paper is as follows. In Sect. 2, we give the related
work. In Sect. 3 is described the distance-based advertisement-delivery system.
In Sect. 4, we provide the description of simulation system and the evaluation
results. Finally, conclusions and future work are given in Sect. 5.
2 Related Work
DTN can provide a reliable internet-working for space tasks [1,3,5,8,15,25].
The space networks have possibly long delay, frequent link disconnection and
frequent disruption. In DTN, the messages are stored and forwarded by nodes.
When the nodes receive messages, they store the messages in their storage. After
that, the nodes duplicate the messages to other nodes when it is transmitted.
This technique is called message switching. The architecture is specified in RFC
4838 [6].
Epidemic routing is well-known routing protocol for DTN [14,16,24]. Epi-
demic routing uses two control messages to duplicate messages. Nodes periodi-
cally broadcast Summary Vector (SV) message in the network. The SV contains
a list of stored messages of each node. When the nodes receive the SV, they com-
pare received SV to their SV. The nodes send REQUEST message if received
SV contains unknown messages.
In Epidemic routing, consumption of network resources and storage usage
become a critical problem. Because the nodes duplicate messages to neighbors
in its communication range. Moreover, received messages remain in the stor-
age, because the messages are continuously duplicated even if destination node
receives the messages. Therefore, recovery schemes such as timer or anti-packet
are needed to limit the duplicate messages. In case of the timer, messages have
lifetime. The messages are punctually deleted when the lifetime of the mes-
sages is expired. However, setting of a suitable lifetime is difficult. In case of
anti-packet, destination node broadcasts the anti-packet, which contains the list
of messages that are received by the destination node. Nodes delete the mes-
sages according to the anti-packet. Then, the nodes duplicate the anti-packet to
other nodes. However, network resources are consumed by the anti-packet. In
this paper, we propose a distance-based message relaying method to limit the
duplicate messages in Vehicular DTN.
3 Distance-Based Advertisement-Delivery System

In this section, we explain in detail of proposed distance-based advertisement-
delivery system.
A Distance-Based Advertisement-Delivery Method for Vehicular DTN 217
(a) First phase (b) Second phase
Fig. 1. Message flooding
3.1 Advertisement-Delivery System

The proposed system delivers advertisement messages to vehicles within a spec-
ified range by inter-vehicle communication using DTN. Our method is based on
Epidemic routing for DTN communication. Each vehicle has a GPS module con-
nected to their OBU to obtain accurate latitude and longitude. We have added
two types of information (location of message originator’s and message dissemi-
nation scope) in a header of the message used in Epidemic routing. We added a
recovery scheme considering these information to delete the replicated bundles in
storage of each vehicle. The process of message flooding is shown in Fig. 1. In this
case, message originator is car shop to broadcast advertisement. The originator
is also considered for general vehicles. At first phase, when the vehicle receives
a message, it checks its position. If the position of vehicle is inside of service
range, the message is stored in the vehicle. Black and red vehicles duplicate the
message to other vehicles during moving on the road. At second phase, when
the position of green vehicle is out of service range, the message is discarded. In
this way, our method decreases the storage usage and replicated message in the
network.
3.2 Distance-Based Recovery Method

We present the flowchart of the recovery method in Fig. 2. In our recovery
method, we delete the replicated bundle messages in storage of each node based
on service range, such as zone filter. In our method, every vehicle periodically
check the distance in intervals of 1 s.
4 Modeling and Simulation Results

In this paper, we evaluate the proposed recovery method for different vehi-
cles. We implemented the proposed recovery method on Scenargie [18] network
simulator.
(a) Conventional Epidemic (b) Proposed method
Fig. 2. Flowcharts of message exchange.
4.1 Scenario Setting
We use two types of routing protocols and send bundle messages from starting-
point (car shop) to all vehicles considering ITU-R P.1411 propagation model [2].
Therefore, we consider interference of obstacles on 5.9 GHz band.
1. Conventional Epidemic
2. Epidemic with proposed distance-based message relaying method
Fig. 3. Overview of zone filter.

Table 1. Simulation parameters.
Parameter Value
Simulation time (Tmax ) 600 [s]
Area dimensions 2, 000 [m] × 2, 000 [m]
Number of vehicles 100, 200, 300 [vehicles]
Service range 1, 000 [m]
Minimum speed (Vmin ) 8.333 [m/s]
Maximum speed (Vmax ) 16.666 [m/s]
PHY model IEEE 802.11p
Propagation model ITU-R P.1411
Antenna model Omni-directional
We consider a grid scenario with 100, 200 and 300 vehicles. We present the sim-
ulation scenario for 100 vehicles in Fig. 3. Table 1 shows the simulation parame-
ters. Message starting point is static, and other vehicles move on road based on
random way-point mobility model.
For simulations, we consider two evaluation parameters: position distribution
and delay. The position distribution indicates the vehicle positions when vehicle
receives the bundle message in service range. The delay indicates the duration
of the message to reach the end of service range. We show the edge nodes on
simulation area in Fig. 3.
4.2 Simulation Results

We present the simulation results of position distribution for 300 vehicles in
Fig. 4. For conventional Epidemic, we confirmed that the vehicles received the
bundle messages on all roads. The results of position distribution are decreased
Fig. 4. Position distribution, when vehicles receive the bundle messages using conven-
tional Epidemic.
(a) 100 vehicles
(b) 200 vehicles
(c) 300 vehicles
Fig. 5. Position distributions for different vehicles, when vehicles receive the bundle
messages using proposed method. Red line indicates the edge of service range.
Table 2. Results of delay for different vehicles, which are measured by edge of service
range.
Number of vehicles 100 200 300

Delay [sec] 37.811 15.105 8.113
compared with conventional Epidemic. We observed that our proposed method

limits the replicated bundle messages for different vehicles in Fig. 5. The service
range was fixed to 1,000 m. The vehicle position distributions were not accurately
within the specified service range. Because each vehicle can not detect the service
range of the bundle message before it actually receives a message. We can see that
position distribution out of service range is increased with increase of vehicles.
The results of delay are shown in Table 2. The results of delay decreased by
increasing the number of vehicles. For 100 vehicles, the results of mean delay was
about 38 s. If the vehicle speed is 60 km/h, the theoretical distance that vehicle
will travel is about 630 m. We observed that the proposed method is an effective
method even the vehicle moves at high speed.
5 Conclusions
In this paper, we proposed a distance-based recovery method for Vehicular DTN.
We evaluated the proposed recovery method considering position distribution
and delay as evaluation metrics. From the simulation results, we observed that
our proposed method limits the replicated bundle messages considering specified
area by applying the distance-based advertisement-delivery method.
In the future work, we would like to add new functions and make extensive
simulations considering different parameters.
References
1. Delay- and disruption-tolerant networks (DTNs) tutorial. NASA/JPL’s Interplan-
etary Internet (IPN) Project (2012). http://www.warthman.com/images/DTN
Tutorial v2.0.pdf
2. Recommendation ITU-R P.1411-7: Propagation data and prediction methods for
the planning of short-range outdoor radiocommunication systems and radio local
area networks in the frequency range 300 MHz to 100 GHz. ITU (2013)
3. Araniti, G., Bezirgiannidis, N., Birrane, E., Bisio, I., Burleigh, S., Caini, C., Feld-
mann, M., Marchese, M., Segui, J., Suzuki, K.: Contact graph routing in DTN
space networks: overview, enhancements and performance. IEEE Commun. Mag.
53(3), 38–46 (2015)
4. Araniti, G., Campolo, C., Condoluci, M., Iera, A., Molinaro, A.: LTE for vehicular
networking: a survey. IEEE Commun. Mag. 21(5), 148–157 (2013)
5. Caini, C., Cruickshank, H., Farrell, S., Marchese, M.: Delay- and disruption-
tolerant networking (DTN): an alternative solution for future satellite networking
applications. Proc. IEEE 99(11), 1980–1997 (2011)
6. Cerf, V., Burleigh, S., Hooke, A., Torgerson, L., Durst, R., Scott, K., Fall, K., Weiss,
H.: Delay-tolerant networking architecture. IETF RFC 4838 (Informational), April
2007
7. Dias, J.A.F.F., Rodrigues, J.J.P.C., Xia, F., Mavromoustakis, C.X.: A cooperative
watchdog system to detect misbehavior nodes in vehicular delay-tolerant networks.
IEEE Trans. Ind. Electron. 62(12), 7929–7937 (2015)
8. Fall, K.: A delay-tolerant network architecture for challenged Internets. In: Pro-
ceedings of the International Conference on Applications, Technologies, Architec-
tures, and Protocols for Computer Communications, SIGCOMM 2003, pp. 27–34
(2003)
9. Grassi, G., Pesavento, D., Pau, G., Vuyyuru, R., Wakikawa, R., Zhang, L.: VANET
via named data networking. In: Proceedings of the IEEE Conference on Computer
Communications Workshops (INFOCOM WKSHPS 2014), pp. 410–415, April 2014
10. Ikeda, M., Ishikawa, S., Barolli, L.: An enhanced message suppression controller for
vehicular-delay tolerant networks. In: Proceedings of the 30th IEEE International
Conference on Advanced Information Networking and Applications (IEEE AINA-
2016), pp. 573–579, March 2016
11. Mahmoud, A., Noureldin, A., Hassanein, H.S.: VANETs positioning in urban envi-
ronments: a novel cooperative approach. In: Proceedings of the IEEE 82nd Vehic-
ular Technology Conference (VTC-2015 Fall), pp. 1–7, September 2015
12. Ohn-Bar, E., Trivedi, M.M.: Learning to detect vehicles by clustering appearance
patterns. IEEE Trans. Intell. Transp. Syst. 16(5), 2511–2521 (2015)
13. Radenkovic, M., Walker, A.: CognitiveCharge: disconnection tolerant adaptive col-
laborative and predictive vehicular charging. In: Proceedings of the 4th ACM
MobiHoc Workshop on Experiences with the Design and Implementation of Smart
Objects (SMARTOBJECTS-2018), June 2018
14. Ramanathan, R., Hansen, R., Basu, P., Hain, R.R., Krishnan, R.: Prioritized epi-
demic routing for opportunistic networks. In: Proceedings of the 1st International
MobiSys Workshop on Mobile Opportunistic Networking (MobiOpp 2007), pp.
62–66 (2007)
15. Rüsch, S., Schürmann, D., Kapitza, R., Wolf, L.: Forward secure delay-tolerant
networking. In: Proceedings of the 12th Workshop on Challenged Networks
(CHANTS-2017), pp. 7–12, October 2017
16. Rohrer, J.P., Mauldin, A.N.: Implementation of epidemic routing with IP conver-
gence layer in ns-3. In: Proceedings of the 10th Workshop on NS-3 (WNS3-2018),
pp. 69–76, June 2018
17. Sato, F., Kikuchi, R.: Hybrid routing scheme combining with geo-routing and DTN
in VANET. In: Proceedings of the 10th International Conference on Innovative
Mobile and Internet Services in Ubiquitous Computing (IMIS-2016), pp. 250–255,
July 2016
18. Scenargie: Space-time engineering, LLC. http://www.spacetime-eng.com/
19. Stute, M., Maass, M., Schons, T., Hollick, M.: Reverse engineering human mobility
in large-scale natural disasters. In: Proceedings of the 20th ACM International
Conference on Modelling, Analysis and Simulation of Wireless and Mobile Systems
(MSWiM-2017), pp. 219–226, November 2017
20. Theodoropoulos, T., Damousis, Y., Amditis, A.: A load balancing control algorithm
for EV static and dynamic wireless charging. In: Proceedings of the IEEE 81st
Vehicular Technology Conference (VTC-2015 Spring), pp. 1–5, May 2015
21. Tornell, S.M., Calafate, C.T., Cano, J.C., Manzoni, P.: DTN protocols for vehicular
networks: an application oriented overview. IEEE Commun. Surv. Tutor. 17(2),
868–887 (2015)
22. Uchida, N., Ishida, T., Shibata, Y.: Delay tolerant networks-based vehicle-to-
vehicle wireless networks for road surveillance systems in local areas. Int. J. Space
Based Situated Comput. 6(1), 12–20 (2016)
23. Urquiza-Aguiar, L., Igartua, M.A., Tripp-Barba, C., Calderón-Hinojosa, X.:
2hGAR: 2-hops geographical anycast routing protocol for vehicle-to-infrastructure
communications. In: Proceedings of the 15th ACM International Symposium
on Mobility Management and Wireless Access (MobiWac-2017), pp. 145–152,
November 2017
24. Vahdat, A., Becker, D.: Epidemic routing for partially-connected ad hoc networks.
Duke University, Technical report (2000)
25. Wyatt, J., Burleigh, S., Jones, R., Torgerson, L., Wissler, S.: Disruption toler-
ant networking flight validation experiment on NASA’s EPOXI mission. In: Pro-
ceedings of the 1st International Conference on Advances in Satellite and Space
Communications (SPACOMM-2009), pp. 187–196, July 2009
26. Yoshino, Y., Nakasaki, S., Ikeda, M., Barolli, L.: A threshold-based adaptive
method for message suppression controller in vehicular DTNs. In: Proceedings
of the 13th International Conference on Broad-Band Wireless Computing, Com-
munication and Applications (BWCCA-2018), pp. 517–524, October 2018
27. Zhang, W., Jiang, S., Zhu, X., Wang, Y.: Cooperative downloading with privacy
preservation and access control for value-added services in VANETs. Int. J. Grid
Util. Comput. 7(1), 50–60 (2016)
Numerical Simulation of Glide Slope
Signal Interferences by Irregular Ground
Junichi Honda1(B) , Hirohisa Tajima1 , and Hisashi Yokoyama2

1
Electronic Navigation Research Institute (ENRI), National Institute of Maritime,
Port, Aviation Technology, Chofu, Tokyo 182-0012, Japan
j-honda@enri.go.jp
2
Japan Aviation Consultant D&T Co., Ltd. (ACDT),
Niiza, Saitama Prefecture 352-0031, Japan
Abstract. This paper focuses on the performance degradation of glide

slope (GS). GS guides pilots regarding their vertical position with respect
to an approach path that is formed using ground-reflections. Irregular
ground features are known to scatter electromagnetic waves and produce
positional errors. Numerical simulations can provide possible solutions
for predicting unexpected influences of these scattered electromagnetic
waves. We discuss numerical techniques based on the ray-tracing method
for two and three dimensional ground models. Propagation paths are
estimated from the shortest path along a straight line between a source
and a receiver. We discuss how the proposed numerical method works
for different ground shapes.
1 Introduction
Instrument Landing System (ILS) plays an important role in the safe landing
of aircraft at airports. ILS consists of a Localizer (LOC), Glide Slope (GS), and
Maker [1]. The GS provides guidance on the vertical position of the aircraft
with reflect to the approach path to the runway center [1]. The radiation from
GS antennas forms an approach path using ground-reflections. Specification of
GS is evaluated using the difference of the depth of the modulation (DDM).
When the aircraft moves along an appropriate approach path, DDM becomes
zero, resulting in no movement of the cross point indicator [2]. The value of
the DDM is regulated by the International Civil Aviation Organization (ICAO)
[1], and the quality of the GS is classified into three categories (CAT I, CAT
II, and CAT III). The GS specification of each airport environment should be
satisfied with the determined category. However, if there are irregular ground
features around the GS antenna unnecessary scattered waves are formed, which
impact the approach path. Therefore, it is extremely important to investigate
such multipath environment to maintain the quality of the GS.
Presently, there are many studies on ILS signals [3–11]. In most studies,
ground model for numerical simulation was limited to two-dimensional (2D)
situation, which changes only area in front of GS antenna. However, in reality,
https://doi.org/10.1007/978-3-030-15035-8_21
Numerical Simulation of Glide Slope Signal Interferences 225
there are often slopes on the ground, which have a former area and side area,
resulting in a three dimensional (3D) ground situation. As far as we know, current
comparisons between 2D and 3D models have been insufficient. The purpose of
this study is to analyze the multipath interference caused by irregular ground.
A numerical method based on the ray-tracing method (RTM) [12–14] is used
to compute electromagnetic fields above 2D and 3D ground models. The RTM
is an approximate solution that does not need a lot of computational memory.
Therefore, it can treat a large propagation environment. Here, irregular ground
models are numerically generated by Gaussian random valuable installed to the
software. By comparing the results of 2D and 3D ground models, we evaluate
how the proposed numerical method copes with the different ground shapes.
2 Glide Slope
GS guides an aircraft regarding its vertical position with respect to the app-
roach path. Figure 1 illustrates an approach path formed by the radiation of
GS antenna. In Japan, an M-array transmission system is employed using a
frequency band in the range 329–335 MHz, where the transmitted signals are
generated by amplitude modulation. There are three components to GS radia-
tion: carrier, sideband and clearance. The specification of the GS is defined by
the DDM, which is computed by the difference of the depth of the modulation
between 90 and 150 Hz [1]. GS is located as shown in Fig. 1b which becomes
generally the touch down point. In this figure, L is the distance from the end of
the runway, and D is the distance from the center of the runway.
An approach path is generated from direct waves and ground-reflection
waves. Ground reflections are observed from flat ground in the basic concept
of GS. Scattered waves result in the course disturbance when the ground surface
in front of GS antenna is irregular. Figure 2 illustrates an overview of the elec-
tromagnetic wave interference. Each airport manages an operation method that
suites the category (CAT I, II, or III), and the DDM should be satisfied by the
category in operation. To satisfy the operation category, a numerical simulation
would be an effective solution.
3 Numerical Method
Most research papers on GS considered the ground to be a 2D model, which
changes the height of only former region of the antenna [10,11]. However, in
reality, the changes are in the former region and the side region, therefore, 3D
ground model must be assumed. The behaviors of electromagnetic waves above
a 3D ground model might be more complicated than that of a 2D model. Here,
we propose a numerical method for a 3D ground model and compare its results
to those from a 2D model.
In this paper, RTM is introduced to analyze electromagnetic waves above
irregular ground. It can treat a relatively large analytical region than other
methods [15,16]. First, the RTM searches for rays from a source to a receiver.
226 J. Honda et al.
(a) Radiation pattern of GS
(b) Positional relation of the GS to the runway
Fig. 1. Schematic representation of the GS.
Fig. 2. Interference sources of GS signal.
Second, the RTM computes electromagnetic fields based on the ray data. The
computational time required to search for the rays depends on the propagation
environment. In general, the searching rays dominate about 90% of the total
computation time. Since our target is a large analytical region, a simplified RTM
is applied to keep the required computer memory as small as possible [12]. In

this section, the principle of the proposed simplified RTM is reviewed.
(a) 2D ground model in 3D point of view
(b) 3D ground model in view of x-y plane. The height in z-axis is given at each intersection between
triangular faces.
Fig. 3. Schematic representations of 2D and 3D ground models.
3.1 Ground Model

The RTM requires a lot of computer memory if we are to treat a complicated
ground surface. Therefore, we discretize the ground into rectangular plates for
a 2D model and triangular plates for a 3D model. Gaussian random variable
installed to the software was used for generating grounds. The 2D ground model
does not change along the side direction as shown in Fig. 3a. Ray searching from
a source to a receiver is estimated from the position and normal vectors of each
rectangular plate. Conversely, the constitution of the 3D model is much more
complicated than that of the 2D model. The ground is divided into triangular
faces as shown in Fig. 3b, and normal vectors are determined corresponding to
the number of faces. The position vector and the relation between neighboring
faces are stored. Therefore, a larger computational cost is required for a 3D
228 J. Honda et al.
model than a 2D model. However, the 3D ground model is much closer to the
real environment.
3.2 Ray Searching
This study considered the shortest path along a straight line between a source to
a receiver. Any other plate/face that is not in a straight line is ignored to simplify
the computation. The first step is to evaluate the relation between a source, a
receiver, and each face. The center position on each face is then defined. If they
are in line of sight (LOS), an approximate ray is generated from the source to
the receiver via the central position of each face. This should be modified for
real rays that satisfy Snell’s law using the imaging method. Then, all estimated
rays are filtered to avoid the transmission of the rays against other faces. These
modified rays are reflection rays. Conversely, diffracted rays are estimated at all
faces, which become LOS with the source and receiver. Here, we assume that
only diffracted waves along a straight line between the source and the receiver
is dominant in the total field. Actually, all regions of irregular ground cause
diffraction and reflection, but here we ignore these interferences. Figure 4 shows
a schematic of ray searching above a 2D ground model. The basic ray searching
of a 3D ground model is similar to the 2D ground model.
Fig. 4. The principle of search rays that are connected by the shortest path along a
straight line from source to receiver.
3.3 Numerical Method

Based on the ray information, the electric field is computed. Let E0u,m,l be the
electric field of each antenna. The received power is defined by:
⎡ i s
⎤
N m=Mn k=Mn
e−jκrn
u,m,l
Eru,m,l
= ⎣ i
(Dnm ) · s u,m,l
(Dnk ) · E0 ⎦ (1)
n=1 m=1 k=1 rnu,m,l
where superscripts u, m, and l represent the upper, middle, and lower antennas,
respectively. N is the total number of rays, Mni is the number of times of image
diffraction, and Mns is the number of times of source diffraction. Ds and Di are
the diffraction coefficients for the source and image, respectively. The reflection
coefficient is included in the image diffraction. The distance of the n-th ray from
the source to the receiver is given by:
i s
k=Mn
+Mn
rnu,m,l = rnk (n = 1, 2, . . . , N ) (2)
k=0
where rnk is the k-th distance from one reflection or diffraction point to the
next one. Diffraction is divided into two parts, i.e., source diffraction and image
diffraction [15], and they work to keep the field continuity as discussed in detail
elsewhere [16].
4 Numerical Result
Figure 5 shows the generated 2D and 3D ground models. Ground shapes were
give by Gaussian random valuable installing in the software. Maximum height
is about 2.5 m, and minimum height is about −3.0 m. The ground in the 2D
model changes in the direction of the forward area, whereas the ground in the
3D model is generated from the forward area and side area. Therefore, the 3D
ground model is more complicated than the 2D model. The GS antenna is located
at L = 315 m and D = 120 m.
Figure 6 shows some examples of search rays, including the first reflection and
diffraction. The receiver is located at x = 1000 m, and the height is computed
to be 3.0◦ of approach path. In Fig. 6a, direct rays and diffracted rays caused by
the edge of the ground appear. Rays are emitted from three antennas installed
to the GS. In Fig. 6b, direct, reflected, and diffracted rays are computed. It is
shown that the propagation path above 3D ground is more complicated than
that above 2D ground. In the 3D case, the rays are estimated along a straight
line between the source and the receiver where many intersections exist.
Based on previous ray information, we can compute the DDMs for 2D and
3D models. Figure 7 shows a comparison of DDMs, which are computed above
flat, 2D, and 3D grounds. The red line is the DDM above flat ground. This
expresses the ideal condition where no interference is detected. The green line
shows the DDM above 2D ground, and some small interference is observed, which
230 J. Honda et al.
(a) 2D ground model
(b) 3D ground model
Fig. 5. 2D and 3D generated ground models.
gradually increases when the receiver is far from the GS antenna. This might be
caused by an average slope in front of the antenna, and is regarded as a primary
plane. The blue line shows the DDM above 3D ground. A lot of interference is
observed because the ground is more complicated than that in the 2D model.
There is a discontinuous section between 3500 m and 6200 m. In this region, a
reflected wave that satisfied Snell’s law is computed, but reflections in the other
regions cannot be detected, except for the near field. In this simulation, ground
(a) Rays above the 2D ground model
(b) Rays above the 3D ground model
Fig. 6. Example of searched rays.
shapes of 2D and 3D in front of the GS antenna are absolutely identical with

each other. However, the 3D ground also changes in the side direction causing
a lot of interference. The discontinuous section could be reduced by including
higher orders of scattered waves, but this paper only considered the first order
of scattered waves.
232 J. Honda et al.
Fig. 7. Comparison of DDMs above flat, 2D, and 3D ground models.
5 Conclusion
This paper discussed the interference of electromagnetic waves above irregular

ground that affects GS signals. Here, we compared 3D ground models with 2D
ground models. To reduce computational resources, 3D ground was generated
using triangular faces. Additionally, only rays that made a straight line between
the source and the receiver were considered. Diffraction was computed by the
intersections between each edge of a discretized face and the straight line from
the source to the receiver. The number of faces in the 3D model is larger than
that in the 2D model, thus the number of rays were larger. Moreover, it was found
that the DDM in the 3D model was fluctuated more than in the 2D model.
Here, we only introduced a concept for analyzing GS signals above a 3D
ground model. The numerical accuracy of the model should be validated and
the proposed method should be compared with the results from other numerical
approaches. This will be considered in our future work.
Acknowledgments. The work was supported in part by a Grant-in-Aid for Young

Scientists (B) (16K18072) from Japan Society for Promotion of Science.
References
1. ICAO (International Civil Aviation Organization): Aeronautical Telecommunica-
tions, Annex 10, vol. 1, 6th edn. (2006)
2. Chin, G., Jordan, L., Kahn, D., Morin, S.: Instrument landing system performance
prediction. IEEE MTT-S International Microwave Symposium, pp. 346–348, May
1975
3. Redlich, R.W., Gorman, J.T.: Disturbance of ILS localizer signals by reflections
from large hangers. IEEE Trans. Aerosp. Electron. Sys. AES-5, 1001–1002 (1969)
4. Shin, S.L.: ILS localizer multipath analysis. IEEE Trans. Aerosp. Electron. Sys.
AES-7(1), 54–60 (1971)
5. Thain, A., Estienne, J.-P., Robert, J., Peres, G., Cambon, G., Evain, L., Splitz, B.:
A solution for ILS disturbance due to a building. In: Proceedings of the 6th Euro-
pean Conference on Antenna and Propagation (EuCAP), pp. 2392–2395, March
2012
6. Thain, A., Estienne, J.P., Peres, G., Spitz, B., Evain, L., Kumar, M.: Comparisons
of different approaches for ILS simulation. In: Proceedings of the 4th European
Conference on Antenna and Propagation (EuCAP), Barocelona, Spain, pp. 1–5,
April 2010
7. Geise, R., Schueuer, J., Thiele, L., Notte, K., Beckers, T., Enders, A.: A slotted
waveguide setup as scaled instrument-landing-system for measuring scattering of
an A380 and large objects. In: Proceedings of the Fourth European Conference on
Antennas and Propagation (EuCAP 2010), Barcelona, Spain, pp. 227–231, April
2010
8. Greving, G., Biermann, W.D.: Theoretical aspect and numerical solution of scat-
tered from objects under grazing angle incidence - the A380 and the Aircraft Tailfin.
In: Proceedings of the 37th European Microwave Conference, Munich, Germany,
pp. 1369–1372, October 2007
9. Godfrey, J., et al.: Terrain modeling using the half-plane geometry with appli-
cations to ILS glide slope antenna. IEEE Trans. Antennas Propag. 24, 370–378
(1976)
10. Luebbers, R., Ungvichian, V., Mitchell, L.: GTD terrain reflection model applied
to ILS glide scope. IEEE Trans. Aerosp. Electron. Sys. AES-18, 11–20 (1982)
11. Walton, E.K.: Effect of wet snow on the null-reference ILS system. IEEE Trans.
Aerosp. Electron. Syst. 29, 1030–1035 (1993)
12. Honda, J., Otsuyama, T.: An estimation algorithm of scattered powers caused by
a moving aircraft. IEICE Commun. Express 2(11), 490–495 (2013)
13. Honda, J., Otsuyama, T.: Estimation of target detection rate in aircraft surveillance
system. In: Proceedings of the 17th International Conference on Network-Based
Information Systems (NBiS 2014), Salerno, Italy, pp. 561–565, September 2014
14. Honda, J., Otsuyama, T.: Rapid computation algorithm for radio wave propaga-
tion characteristics on airport surface. In: Proceedings of the Eighth International
Conference on Complex Intelligent and Software Intensive Systems (CISIS 2014),
Birmingham, United Kingdom, pp. 302–306, July 2014
15. Honda, J., Uchida, K., Yoon, K.Y.: Estimation of radio communication distance
along random rough surface. IEICE Trans. ELECTRON. E93-C(1), 39–45 (2010)
16. Takematsu, M., Uchida, K., Honda, J.: Method for reduction of field computation
time for discrete ray tracing method. IEICE Trans. Electron. E97-C(3), 198–206
(2014)
Effect of Client Priority
in the Performance of a Fuzzy-Based
WLAN Triage System
Kosuke Ozera1(B) , Yi Liu1 , Leonard Barolli2 , and Makoto Takizawa3

1
kosuke.o.fit@gmail.com, ryuui1010@gmail.com
2
Fukuoka Institute of Technology (FIT), 3-30-1 Wajiro-Higashi, Higashi-Ku,
Fukuoka 811-0295, Japan
barolli@fit.ac.jp
3
Department of Advanced Sciences, Faculty of Science and Engineering,
Hosei University, Kajino-Machi, Koganei-Shi, Tokyo 184-8584, Japan
Abstract. Many devices communicate over Wireless Local Area Net-

works (WLANs). The IEEE 802.11e standard for WLANs is an impor-
tant extension of the IEEE 802.11 standard focusing on QoS that works
with any PHY implementation. The IEEE 802.11e standard introduces
EDCF and HCCA. Both these schemes are useful for QoS provisioning
to support delay-sensitive voice and video applications. EDCF uses the
contention window to differentiate high priority and low priority services.
However, it does not consider the priority of users. In order to deal with
this problem, in our previous work, we proposed a Fuzzy-based Admis-
sion Control System (FACS). In this paper, we present the effect of client
priority in the performance of a Fuzzy-based WLAN Triage system. We
concluded that when CKT, RSSI and UP values increase and NP value
decreases, the clients have high priority, so they will keep connection to
the AP.
1 Introduction
With the development of wireless technology and Internet, there is an increasing
need towards portable and mobile computers such as smart phones [1,24]. The
wireless networks need to provide communications between mobile terminals.
The Wireless Local Area Networks (WLANs) provide high bandwidth access for
users in a limited geographical area. With the popularization of mobile devices,
many device communicate together over WLANs [27]. WLANs have a lot of
restriction on communication resources comparing wired LANs. Therefore, it is
more difficult to guarantee the Quality of Service (QoS).
In our previous work, the admission decision is done by a Fuzzy-based Admis-
sion Control System (FACS) [7–9]. In this paper, we present the effect of client
https://doi.org/10.1007/978-3-030-15035-8_22
Effect of Client Priority 235
Fig. 1. CSMA/CA.
priority in the performance of a Fuzzy-based WLAN Triage system. We con-

cluded that when CKT, RSSI and UP values increase and NP value decreases,
the clients have high pirority, so they will keep connection to the AP.
The paper is organized as follows. In Sect. 2, we give introduce of CSMA/CA.
In Sect. 3, we discuss the application of Fuzzy Logic (FL) for control. In Sect. 4,
we present the proposed system. In Sect. 5, we show the simulation results.
Finally, in Sect. 6, we conclude the paper.
2 CSMA/CA
WLAN is standardized by IEEE 802.11 which uses CSMA/CA as shown in
Fig. 1. The nodes check whether other devices are communicating or not before
starting communication [5]. If other nodes are communicating with the AP, the
node waits for a period of time, called Distributed Inter Frame Space (DIFS).
After that, it waits an additionally random time called back-off time. After the
back-off time, if other devices are not communicating, the node starts to send
data [25]. AP which received the data waits a constant time, called Short Inter
Frame Space (SIFS) and sends ACK to the node which sent the data. Because
ACK frame should be sent soon, it is shorter than DIFS.
The IEEE 802.11e standard is an important extension of the IEEE 802.11
standard focusing on QoS [16] that works with any PHY implementation. Wire-
less nodes equipped with IEEE 802.11e features are now known as QoS stations
(QSTAs) and they are associated with a QoS access point (QAP) to form a QoS
basic service set (QBSS). The main feature of the IEEE 802.11e standard is
that it improves the MAC layer for QoS provisioning by providing support for:
segregation of data packets based on priority requirements; negotiation of QoS
parameters through a central coordinator or AP; and admission control.
The IEEE 802.11e standard introduces EDCF. This scheme is useful for QoS
provisioning to support delay-sensitive voice and video applications [2].
In the DCF configuration, a contention window is set after a frame is trans-
mitted. This is done to avoid any collisions. The window defines the contention
time of various stations who contend with each other for access to channel. How-
ever, each of the stations cannot size the channel immediately, rather the MAC
protocol uses a randomly chosen time period for each station after that channel
has undergone transmission [28].
236 K. Ozera et al.
The EDCF uses this contention window to differentiate between high prior-
ity and low priority services [23]. The central coordinator assigns a contention
window of shorter length to the stations with higher priority that helps them to
transmit before the lower priority ones [22]. To differentiate further, Inter Frame
Spacing (IFS) can be varied according to different traffic categories. Instead of
using a DIFS as for the DCF traffic, a new inter-frame spacing called Arbitration
Inter Frame Spacing (AIFS) is used. The AIFS used for traffic has a duration
of a few time slots longer than the DIFS duration. Therefore, a traffic category
having smaller AIFS gets higher priority.
IEEE802.11e provides EDCF and Hybrid coordination function controlled
channel access (HCCA) as a priority control method [2,17,30]. Mainly, EDCF
is used because of easy implementation and compatibility of CSMA/CA.
3 Application of Fuzzy Logic for Control

The ability of fuzzy sets and possibility theory to model gradual properties or
soft constraints whose satisfaction is matter of degree, as well as information
pervaded with imprecision and uncertainty, makes them useful in a great variety
of applications.
The most popular area of application is Fuzzy Control (FC), since the appear-
ance, especially in Japan, of industrial applications in domestic appliances,
process control, and automotive systems, among many other fields [3,4,6,12–
15,18,26].
3.1 FC
In the FC systems, expert knowledge is encoded in the form of fuzzy rules, which
describe recommended actions for different classes of situations represented by
fuzzy sets.
In fact, any kind of control law can be modeled by the FC methodology,
provided that this law is expressible in terms of “if ... then ...” rules, just like
in the case of expert systems. However, FL diverges from the standard expert
system approach by providing an interpolation mechanism from several rules. In
the contents of complex processes, it may turn out to be more practical to get
knowledge from an expert operator than to calculate an optimal control, due to
modeling costs or because a model is out of reach.
3.2 Linguistic Variables
A concept that plays a central role in the application of FL is that of a linguistic

variable. The linguistic variables may be viewed as a form of data compression.
One linguistic variable may represent many numerical variables. It is suggestive
to refer to this form of data compression as granulation [10].
The same effect can be achieved by conventional quantization, but in the

case of quantization, the values are intervals, whereas in the case of granula-
tion the values are overlapping fuzzy sets. The advantages of granulation over
quantization are as follows:
• it is more general;
• it mimics the way in which humans interpret linguistic values;
• the transition from one linguistic value to a contiguous linguistic value is
gradual rather than abrupt, resulting in continuity and robustness.
3.3 FC Rules
FC describes the algorithm for process control as a fuzzy relation between infor-
mation about the conditions of the process to be controlled, x and y, and the
output for the process z. The control algorithm is given in “if ... then ...” expres-
sion, such as:
If x is small and y is big, then z is medium;
If x is big and y is medium, then z is big.
These rules are called FC rules. The “if” clause of the rules is called the
antecedent and the “then” clause is called consequent. In general, variables x
and y are called the input and z the output. The “small” and “big” are fuzzy
values for x and y, and they are expressed by fuzzy sets.
Fuzzy controllers are constructed of groups of these FC rules, and when an
actual input is given, the output is calculated by means of fuzzy inference.
3.4 Control Knowledge Base

There are two main tasks in designing the control knowledge base. First, a set
of linguistic variables must be selected which describe the values of the main
control parameters of the process. Both the input and output parameters must
be linguistically defined in this stage using proper term sets. The selection of the
level of granularity of a term set for an input variable or an output variable plays
an important role in the smoothness of control. Second, a control knowledge base
must be developed which uses the above linguistic description of the input and
output parameters. Four methods [19,21,29,31] have been suggested for doing
this:
• expert’s experience and knowledge;
• modelling the operator’s control action;
• modelling a process;
• self organization.
Among the above methods, the first one is the most widely used. In the
modeling of the human expert operator’s knowledge, fuzzy rules of the form
“If Error is small and Change-in-error is small then the Force is small” have
been used in several studies [11,20]. This method is effective when expert human
operators can express the heuristics or the knowledge that they use in controlling
a process in terms of rules of the above form.
238 K. Ozera et al.
Fig. 2. Proposed system model.
3.5 Defuzzification Methods
The defuzzification operation produces a non-FC action that best represent the
membership function of an inferred FC action. Several defuzzification methods
have been suggested in literature. Among them, four methods which have been
applied most often are:
• Tsukamoto’s Defuzzification Method;

• The Center of Area (COA) Method;
• The Mean of Maximum (MOM) Method;
• Defuzzification when Output of Rules are Function of Their Inputs.
4 Proposed System
The proposed system model is show in Fig. 2. For implementing our proposed sys-
tem, we consider 4 linguistic input parameters: Connected Keeping Time (CKT),
User Priority (UP), Received Signal Strength Indication (RSSI) and Number of
Packet (NP) to decide the Connection Priority (CP) output parameter. The mem-
bership functions of proposal system are shown in Fig. 3. In Table 1, we show the
Fuzzy Rule Base (FRB) of the proposal system, which consists of 81 rules.
The term sets of CKT, UP, RSSI and NP are defined respectively as:
CKT = {Short, M iddle, Long}
= {S, M, L};
U P = {Low, M iddle, High}
= {Lo, M i, Hi};
RSSI = {Low, M iddle, High}
= {LO, M I, HI}.
N P = {F ew, M iddle, M any}
= {F e, M id, M a};
S M H Lo Mi Hi
1 1
µ(CKT)
µ(UP)
0.5 0.5
0 0
0 900 1800 2700 3600 0 20 40 60 80 100
CKT UP
(a) Connected Keeping Time (b) User Priority
LO MI HI Fe Mid Ma
1 1
µ(RSSI)
µ(NP)
0.5 0.5
0 0
-100 -80 -60 -40 -20 0 0 50 100 150 200 250 300
RSSI NP
(c) Received Signal Strength Indication (d) Number of Packet
EL VVL VL Lo Mid Hi VH VVH EH

1.0
µ(CP)
0.5
0.0
0 0.2 0.4 0.6 0.8 1
CP
(e) Connection Priority
Fig. 3. Membership functions.
and the term set for the output CP is defined as:

⎛ ⎞ ⎛ ⎞
ExtremeLow EL
⎜ V eryV eryLow ⎟ ⎜ V V L ⎟
⎜ ⎟ ⎜ ⎟
⎜ V eryLow ⎟ ⎜ VL ⎟
⎜ ⎟ ⎜ ⎟
⎜ Low ⎟ ⎜ Lo ⎟
⎜ ⎟ ⎜ ⎟
CP = ⎜ M iddle ⎟ = ⎜ M id ⎟ .
⎜ ⎟ ⎜ ⎟
⎜ High ⎟ ⎜ Hi ⎟
⎜ ⎟ ⎜ ⎟
⎜ V eryHigh ⎟ ⎜ V H ⎟
⎜ ⎟ ⎜ ⎟
⎝ V eryV eryHigh ⎠ ⎝ V V H ⎠
ExtremeHigh EH
240 K. Ozera et al.
In this section, we present the simulation results for the proposed system. In our
system, we decided the membership functions by carrying out many simulations.
We show the simulation results in Figs. 4 and 5. We show the relation between
CP and CKT, UP, RSSI and NP. We consider the CKT and NP as a constant
parameters.
In Fig. 4, we consider the CKT value 360. We change the UP value 0 to 100.
In Fig. 4(a), when the UP is increased, the CP is increased. Also, by increasing
RSSI value, the CP is increased. In Fig. 4(b) and (c), we increased the value of
NP to 150 and 270, respectively. We can see that when NP is increased, the CP
is decreased. In Fig. 5, we increase the CKT value to 3240. We can see that by
increasing CKT value, the CP is increased.
We concluded that when CKT, RSSI and UP values increase and NP value
decreases, the clients have high pirority, so they will keep connection to the AP.
Table 1. Fuzzy Rule Base
Rule CKT UP RSSI NP CP Rule CKT UP RSSI NP CP Rule CKT UP RSSI NP CP

1 S Lo LO Fe VL 28 M Lo LO Fe Lo 55 L Lo LO Fe Mid
2 S Lo LO Mid Mid 29 M Lo LO Mid Hi 56 L Lo LO Mid Hi
3 S Lo LO Ma VH 30 M Lo LO Ma VH 57 L Lo LO Ma VVH
4 S Lo MI Fe EL 31 M Lo MI Fe VVL 58 L Lo MI Fe VVL
5 S Lo MI Mid VL 32 M Lo MI Mid VL 59 L Lo MI Mid Lo
6 S Lo MI Ma Mid 33 M Lo MI Ma Mid 60 L Lo MI Ma Hi
7 S Lo HI Fe EL 34 M Lo HI Fe EL 61 L Lo HI Fe EL
8 S Lo HI Mid EL 35 M Lo HI Mid EL 62 L Lo HI Mid VVL
9 S Lo HI Ma VVL 36 M Lo HI Ma VL 63 L Lo HI Ma VL
10 S Mi LO Fe Lo 37 M Mi LO Fe Lo 64 L Mi LO Fe Mid
11 S Mi LO Mid Hi 38 M Mi LO Mid Hi 65 L Mi LO Mid VH
12 S Mi LO Ma VH 39 M Mi LO Ma VVH 66 L Mi LO Ma VVH
13 S Mi MI Fe EL 40 M Mi MI Fe VVL 67 L Mi MI Fe VVL
14 S Mi MI Mid VL 41 M Mi MI Mid Lo 68 L Mi MI Mid Lo
15 S Mi MI Ma Mid 42 M Mi MI Ma Hi 69 L Mi MI Ma Hi
16 S Mi HI Fe EL 43 M Mi HI Fe EL 70 L Mi HI Fe EL
17 S Mi HI Mid EL 44 M Mi HI Mid EL 71 L Mi HI Mid VVL
18 S Mi HI Ma VVL 45 M Mi HI Ma VL 72 L Mi HI Ma Lo
19 S Hi LO Fe Hi 46 M Hi LO Fe VH 73 L Hi LO Fe VH
20 S Hi LO Mid VVH 47 M Hi LO Mid VVH 74 L Hi LO Mid EH
21 S Hi LO Ma EH 48 M Hi LO Ma EH 75 L Hi LO Ma EH
22 S Hi MI Fe L 49 M Hi MI Fe Mid 76 L Hi MI Fe Mid
23 S Hi MI Mid Hi 50 M Hi MI Mid Hi 77 L Hi MI Mid VH
24 S Hi MI Ma VH 51 M Hi MI Ma VVH 78 L Hi MI Ma VVH
25 S Hi HI Fe VVL 52 M Hi HI Fe VVL 79 L Hi HI Fe VL
26 S Hi HI Mid VL 53 M Hi HI Mid Lo 80 L Hi HI Mid Mid
27 S Hi HI Ma Mid 54 M Hi HI Ma Hi 81 L Hi HI Ma Hi
CKT=360 NP=30 CKT=360 NP=150

1 1
RSSI=-10 RSSI=-10
0.9 RSSI=-50 0.9 RSSI=-50
RSSI=-90 RSSI=-90
0.8 0.8
0.7 0.7
0.6 0.6
CP [unit]
CP [unit]
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0 0
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
UP [unit] UP [unit]
(a) CKT=360 NP=30 (b) CKT=360 NP=150
CKT=360 NP=270
1
RSSI=-10
0.9 RSSI=-50
RSSI=-90
0.8
0.7
0.6
CP [unit]
0.5
0.4
0.3
0.2
0.1
0
0 10 20 30 40 50 60 70 80 90 100
UP [unit]
(c) CKT=360 NP=270
Fig. 4. Simulation results when CKT is 360.

CKT=3240 NP=30 CKT=3240 NP=150
1 1
RSSI=-10
0.9 0.9 RSSI=-50
RSSI=-90
0.8 0.8
0.7 0.7
0.6 0.6
CP [unit]
CP [unit]
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
RSSI=-10
0.1 RSSI=-50 0.1
RSSI=-90
0 0
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
UP [unit] UP [unit]
(a) CKT=3240 NP=30 (b) CKT=3240 NP=150
CKT=3240 NP=270
1
RSSI=-10
0.9 RSSI=-50
RSSI=-90
0.8
0.7
0.6
CP [unit]
0.5
0.4
0.3
0.2
0.1
0
0 10 20 30 40 50 60 70 80 90 100
UP [unit]
(c) CKT=3240 NP=270
Fig. 5. Simulation results when CKT is 3240.

242 K. Ozera et al.
6 Conclusions
In this paper, we presented the effect of client priority in the performance of a
Fuzzy-based WLAN Triage system. We considered four parameters: CKT, UP,
RSSI and NP. We evaluated the systems by simulations. We concluded that
when CKT, RSSI and UP values increase and NP value decreases, the clients
have high pirority, so they will keep connection to the AP. In our future work,
we would like to carry out extensive experiments.
References
1. A particle swarm optimisation algorithm to generate inhomogeneous triangular
cells for allocating base stations in urban and suburban areas
2. Choi, S., Del Prado, J., Mangold, S., et al.: IEEE 802.11e contention-based channel
access (EDCF) performance evaluation. In: International Conference on Commu-
nications (ICC 2003), vol. 2, pp. 1151–1156 (2003)
3. Elmazi, D., Kulla, E., Oda, T., Spaho, E., Sakamoto, S., Barolli, L.: A comparison
study of two fuzzy-based systems for selection of actor node in wireless sensor actor
networks. J. Ambient. Intell. Hum. Comput. 6(5), 635–645 (2015)
4. Elmazi, D., Sakamoto, S., Oda, T., Kulla, E., Spaho, E., Barolli, L.: Two fuzzy-
based systems for selection of actor nodes inwireless sensor and actor networks: a
comparison study considering security parameter effect. Mob. Netw. Appl. 21(1),
53–64 (2016)
5. Gao, D., Cai, J., Ngan, K.N.: Admission control in IEEE 802.11e wireless LANs.
IEEE Netw. 19(4), 6–13 (2005)
6. Inaba, T., Sakamoto, S., Oda, T., Ikeda, M., Barolli, L.: A secure-aware call admis-
sion control scheme for wireless cellular networks using fuzzy logic and its perfor-
mance evaluation. J. Mob. Multimed. 11(3&4), 213–222 (2015)
7. Inaba, T., Sakamoto, S., Oda, T., Ikeda, M., Barolli, L.: A QoS-aware admission
control system for WLAN using fuzzy logic. In: The 30th International Confer-
ence on Advanced Information Networking and Applications Workshops (WAINA
2016), pp. 499–505 (2016)
8. Inaba, T., Sakamoto, S., Oda, T., Ikeda, M., Barolli, L.: A testbed for admission
control in WLAN: a fuzzy approach and its performance evaluation. In: The 11th
International Conference on Broadband and Wireless Computing, Communication
and Applications (BWCCA 2016), pp. 559–571 (2016)
9. Javanmardi, S., Shojafar, M., Shariatmadari, S., Ahrabi, S.S.: FRTRUST: a fuzzy
reputation-based model for trust management in semantic P2P grids. Int. J. Grid
Util. Comput. 6(1), 57–66 (2014)
10. Kandel, A.: Fuzzy Expert Systems. CRC Press, Boca Raton (1991)
11. Klir, G.J., Folger, T.A.: Fuzzy Sets, Uncertainty, and Information. Prentice Hall,
Englewood Cliffs (1988)
12. Kolici, V., Inaba, T., Lala, A., Mino, G., Sakamoto, S., Barolli, L.: A fuzzy-based
CAC scheme for cellular networks considering security. In: 2014 17th International
Conference on Network-Based Information Systems, pp. 368–373. IEEE (2014)
13. Liu, Y., Sakamoto, S., Matsuo, K., Ikeda, M., Barolli, L., Xhafa, F.: Improvement
of JXTA-overlay P2P platform: evaluation for medical application and reliability.
Int. J. Distrib. Syst. Technol. (IJDST) 6(2), 45–62 (2015)
14. Liu, Y., Sakamoto, S., Matsuo, K., Ikeda, M., Barolli, L., Xhafa, F.: Improving
reliability of JXTA-overlay P2P platform: a comparison study for two fuzzy-based
systems. J. High Speed Netw. 21(1), 27–42 (2015)
15. Liu, Y., Sakamoto, S., Matsuo, K., Ikeda, M., Barolli, L., Xhafa, F.: A comparison
study for two fuzzy-based systems: improving reliability and security of JXTA-
overlay P2P platform. Soft Comput. 20(7), 2677–2687 (2016)
16. Mangold, S., Choi, S., Hiertz, G.R., Klein, O., Walke, B.: Analysis of IEEE 802.11e
for QoS support in wireless LANs. IEEE Wirel. Commun. 10(6), 40–50 (2003)
17. Mangold, S., Choi, S., May, P., Klein, O., Hiertz, G., Stibor, L.: IEEE802.11e
wireless LAN for quality of service. Proc. Eur. Wirel. 2, 32–39 (2002)
18. Matsuo, K., Elmazi, D., Liu, Y., Sakamoto, S., Mino, G., Barolli, L.: FACS-MP: a
fuzzy admission control system with many priorities for wireless cellular networks
and its performance evaluation. J. High Speed Netw. 21(1), 1–14 (2015)
19. McNeill, F.M., Thro, E.: Fuzzy Logic: A Practical Approach. Academic Press,
Boston (1994)
20. Munakata, T., Jani, Y.: Fuzzy systems: an overview. Commun. ACM 37(3), 68–76
(1994)
21. Procyk, T.J., Mamdani, E.H.: A linguistic self-organizing process controller. Auto-
matica 15(1), 15–30 (1979)
22. Qashi, R., Bogdan, M., Hänssgen, K.: Evaluating the QoS of WLANs for the
IEEE802.11 EDCF in real-time applications. In: International Conference on Com-
munications and Information Technology (ICCIT 2011), pp. 32–35 (2011)
23. Romdhani, L., Ni, Q., Turletti, T.: Adaptive EDCF: enhanced service differentia-
tion for IEEE802.11 wireless ad-hoc networks. In: Wireless Communications and
Networking (WCNC 2003), vol. 2, pp. 1373–1378 (2003)
24. Sakamoto, S., Oda, T., Ikeda, M., Barolli, L., Xhafa, F.: Implementation and evalu-
ation of a simulation system based on particle swarm optimisation for node place-
ment problem in wireless mesh networks. Int. J. Commun. Netw. Distrib. Syst.
17(1), 1–13 (2016)
25. Song, N.O., Kwak, B.J., Song, J., Miller, L.E.: Enhancement of IEEE802.11 dis-
tributed coordination function with exponential increase exponential decrease
backoff algorithm. In: The 57th IEEE Semiannual Vehicular Technology Confer-
ence, vol. 4, pp. 2775–2778 (2003)
26. Spaho, E., Sakamoto, S., Barolli, L., Xhafa, F., Ikeda, M.: Trustworthiness in P2P:
performance behaviour of two fuzzy-based systems for JXTA-overlay platform. Soft
Comput. 18(9), 1783–1793 (2014)
27. Wu, H., Peng, Y., Long, K., Cheng, S., Ma, J.: Performance of reliable transport
protocol over IEEE802.11 wireless LAN: analysis and enhancement. In: The 21st
Annual Joint Conference of the IEEE Computer and Communications Societies,
vol. 2, pp. 599–607 (2002)
28. Yang, X., Vaidya, N.H.: Priority scheduling in wireless ad hoc networks. In: Pro-
ceedings of the 3rd ACM International Symposium on Mobile Ad Hoc Networking
& Computing, pp. 71–79 (2002)
29. Zadeh, L.A., Kacprzyk, J.: Fuzzy Logic for the Management of Uncertainty. John
Wiley & Sons, Inc., New York (1992)
30. Zhu, J., Fapojuwo, A.O.: A new call admission control method for providing desired
throughput and delay performance in IEEE802.11e wireless LANs. IEEE Trans.
Wirel. Commun. 6(2), 701–709 (2007)
31. Zimmermann, H.J.: Fuzzy Set Theory and Its Applications. Springer Science &
Business Media, New York (1991)
Energy Saving in HetNet Network
Using eNB Parameters Tuning
Narjes Lassoued1,2(B) , Noureddine Boujnah3,4 , and Ridha Bouallegue1

1
SUP’COM, INNOV’COM Research Lab.,
Higher School of Communication of Tunis, University of Carthage, Tunis, Tunisia
narjes.lassoued@hotmail.com, ridha.bouallegue@supcom.rnu.tn
2
National Engineering School of Gabes, ENIG, Gabes, Tunisia
3
Faculty of Science of Gabes, FSG, Gabes, Tunisia
4
Waterford Institute of Technology, TSSG, Waterford, Ireland
bnoureddine@tssg.org
Abstract. Power consumption represents an exciting subject in the field

of mobile communication networks. In the last decades, the radio access
networks have witnessed a denoting augmentation in power consumption,
and this has indirect drawbacks on both environment and economic sides.
The increase of power consumption conduct to both a substantial growth
in CO2 emissions and on the operational expenditure of network opera-
tors. For this reason, minimizing power consumption in mobile cellular
networks has become mandatory. In this paper, we suggest a power sav-
ing scheme in LTE-Advanced based Heterogeneous Networks (HetNet)
by the mean of switching ON/OFF eNodeB (eNB) and tilt modification
of neighbors.
1 Introduction
Long Term Evolution (LTE), designed by 3GPP, is enhanced by the use of new
access techniques and by the improvement of architecture that combines differ-
ent new services. New methods are included to LTE to migrate to LTE Advanced
such as Coordinated Multipoint (CoMP), Massive MIMO, carrier aggregation
and HetNet. The notion of LTE Advanced based on HetNet aims to ameliorate
spectral efficiency. HetNet is a combination between various cell types: micro-
cells, picocells and femtocells, and separate access technology [1].
In recent years, data traffic in the wireless mobile network has gained consid-
erable growth. Consequently, energy consumption of Information and Communi-
cation Technology (ICT) is rising with a staggering rate. This significant increase
in ICT energy consumption has many environmental and economic effects. Thus,
on the ecological side, ICT is responsible practically for 2% of CO2 emissions [2].
Moreover, on the economic side, the percentage of ICT energy consumption
varies from 2% to 10% of the annual worldwide energy consumption [3].
Referring to Fig. 1 representing the consumption of power in a typical wireless
mobile network [4], it is clear that the base station (BS) consume the biggest
https://doi.org/10.1007/978-3-030-15035-8_23
Energy Saving in HetNet Network Using eNB Parameters Tuning 245
Fig. 1. Power consumption of a typical wireless cellular network
part of the energy that can attain 57 % among all networks [4]. Therefore,
minimize energy consumption in the total network amounts to minimizing energy
consumption in the BS.
Several researches are proposed to reduce energy consumption in cellular
networks, the most well-known are the Switch ON/OFF algorithms. The condi-
tion to apply the Switch ON/OFF procedure is different from one technique to
another. The main aim is to reduce the energy consumption by optimizing the
use of the power of BSs with guaranteed QoS.
In [5], according to the traffic variation, the authors switch ON/OFF dynam-
ically the BSs by respecting certain conditions of blocking probability in order
to guarantee the QoS. In [6], BSs are switched ON/OFF randomly to minimize
energy consumption in a UMTS cellular network. An improvement of this work
is given in [7] where authors proposed a Switch ON/OFF algorithm using a
uniform and a hierarchical scenario.
Algorithms of Switch ON/OFF are also applied in LTE and LTE-A networks
through some researches. Notably, in [8], Alexandra et al. try to come up with
an optimum combination of switching ON/OFF eNBs in order to guarantee the
maximum of energy saving. In [9], Narjes et al. implemented a Switch ON/OFF
scheme in LTE network especially during low traffic period and QoS is guaran-
teed using a power up/down procedure. An improvement of this work is detailed
in [10], where the algorithm of switch ON/OFF is implemented in an LTE-A
based HetNet. In this work, the switch ON/OFF procedure is performed based
on the SINR values of UEs. QoS and coverage are maintained using femtocells
and CoMP techniques.
In all works mentioned above, researchers are concentrated on power con-
sumption issue in UMTS, LTE and LTE-A networks. They all use the Switch
ON/OFF procedure that aims to raise the power saving by minimizing the num-
ber of active eNBs during low traffic periods.
In this paper, we are interested by applying algorithms of Switch ON/OFF in
LTE-A based HetNet. Our contribution is to switch off an eNB when the number
of their attached UEs is lower than a given threshold. UEs of Switched off eNB
are assigned after the Switch OFF depending on their measured SINR (Signal
246 N. Lassoued et al.
to Interference plus Noise Ratio) value. Femtocells will serve UEs with good
SINR value and neighbors eNBs will serve UEs with bad SINR value. Coverage
is maintained by applying a tilt variation procedure for the neighboring eNB.
The rest of the paper is organized as follows: Sect. 2, describes the Net-
work architecture. Section 3, introduces the eNodeB power model. In Sect. 4, we
describe the proposed algorithm. Results of simulation are the subject of Sects. 5
and 6 concludes the paper.
2 Network Architecture
For the network architecture, we propose a HetNet dense urban network, in
which, neighboring cells overlaps. The considered topology is formed by N = 7
eNBs. Each eNB can serve one macrocell surrounded by two femtocells as shown
in Fig. 2. For UEs generation, we apply a non uniform distribution of users where
UEs are generated randomly between cells.
Fig. 2. Regular network layout
In our topology, we focused on two main concepts used in LTE-Advanced.
2.1 Femtocells
Due to the use of HetNet, we benefit from the use of femtocells to reduce power
consumption. At first, femtocells are developed to raise capacity and guarantee
coverage for highly densified areas [1]. Moreover, Femtocells are small, inexpen-
sive and are characterized by low energy consumption of BSs [11]. For this reason,
in our work, femtocells are disposed around the eNB to compensate switched off
procedure.
2.2 Antenna Tilt

The tilt represent the inclination or angle of the antenna to its axis. The tilt
is used when we want to reduce interference and/or coverage of some specific
areas [12]. In our case we will make tilt variation procedure in order to offer
coverage to users of Switched OFF eNBs (Fig. 3).
Fig. 3. Tilt presentation
3 Power Calculation Model

We will present, in this section, the power model used to calculate power saving.
Starting by defining the eNB power model. At an instant t, the power of eNB i,
Pi is given by:
Pi (t) = POP (i) + PT X (i).Si (t) (1)
Where, N is the total number of eNBs, PT X is the total transmission power
of eNB i, Si (t) represent the eNB i state at an instant t and POP (i) is the eNB
i operating power given by:
POP (i) = Pa (i) + Ps (i) + Pc (i) (2)
Where, Pa (i) is the eNB i power amplifier including Feeder, Ps (i) is the power
supply for eNB i and Pc (i) is the cooling power consumed by eNB i. The trans-
mitted power PT X (i) is:
Nu

PT X (i) = PT X (j) (3)
j
Where, Nu is the number of eNB i UEs and PT X (j) is the power transmitted
by UE j given by:
PT X (j) = NRB (j).PRB (i) (4)
NRB (j) is the number of resource blocks attributed to UE j and PRB (i) is the
power transmitted per resource block for eNB i. PRB (i) is calculated using the
following equation:
PT X (i)
PRB (i) = (5)
NRB (i)
Where, PT X (i) is the transmitted power defined in Eq. 3 and NRB (i) is the total
number of resource blocks attributed to eNB i. The state of eNB Si (t) is defined
as follows:
1 if ON
Si (t) = (6)
0 if OF F
The total power of the network Ptotal at an instant t is given based on Eq. 1
by:
N
N

Ptotal (t) = [Pi (t)] + ΔP = [POP (i) + PT X (i).Si (t)] + ΔP (7)
i i
Where, Pi is the eNB i power defined in Eq. 1 and ΔP is the amount of added
power after applying the Switch ON/OFF algorithm.
The received power at UE’s side is calculated using the pathloss model defined
in [13] as:
PRX (θ) = PT X .G(θ).D−α (8)
Where, PT X is the eNB transmission power described in Eq. 4. D is the distance
in km between the eNB and UE, α is the propagation exponent selected depend-
ing on the type of propagation medium and G(θ) is the antenna gain defined
based on equations in [12] by:
arctan (D/h)−θ 2
G(θ) = Gmax .10−12.( θ0 )
(9)
Where, Gmax is the maximum antenna gain, D is the distance between

UE and eNB, h is the antenna height, θ0 is the elevation 3dB value, it
◦
may be assumed to be 15 and θ is the angle tilt: θ ∈ [arctan (D/h) −
θ0 . Am /12, .., arctan (D/h) + θ0 . Am /12], where Am = 20dB is the maximum
attenuation. Let Pmin be the power sensitivity of UE, the received power PRX
must be superior than Pmin .
4 Proposed Power Saving Algorithm
Our energy saving algorithm is based on the concept of Switch ON/OFF espe-
cially during low traffic periods when eNBs are little used. We add to this concept
some new features that significantly reduce energy consumption in the network.
Our algorithm behaves as follow: At first, all macro eNBs are activated, and
all femtocells are deactivated. When the number of active UEs attributed to an
eNB i is smaller than a given threshold T, a Switch OFF procedure is applied to
eNB i, at the same time, femtocells existing in the cell range of switched off eNB
are activated to serve UEs surrounding the deactivated eNB. For UEs located
at the cell extremity, they are allowed to the nearest active eNB neighbor.
Switching ON/OFF eNBs achieve a good amount of power saving but it
affects the QoS by reducing the coverage; Thus, when an eNB is switched off, the
coverage will be decreased and thereby, some UEs will be out of cell’s range. For
this reason, we must compensate the coverage area to provide network coverage
to UEs after switching off process. Thus, to achieve coverage continuity we will
perform a power adjustment procedure combined with a tilt variation technique
of the antenna tilt of neighbor eNBs that will serve the attached UE of the
deactivated eNB. The Switch ON/OFF procedure is stopped when the eNB that
we decide to switch OFF contains some active UEs existing in the cell extremity
and there is no neighbor eNB in ON state. Steps for our proposed approach are:
• Step 1: At first, we start by calculating the number of active UEs Ni (t)

attributed to eNB i.
• Step 2: The second step is to compare the value of calculated number Ni (t) to
a fixed threshold T . We can meet two possible cases described in the algorithm
below:
Algorithm 1. Decision of Switch ON/OFF

if Ni (t) > T then
if Si (t − 1) = 1 then
Si (t) ← 1
else if Si (t − 1) = 0 then
Si (t) ← 1
end if
else if Ni (t) ≤ T then
Si (t) ← 1
Sfi (t) ← 1 {Sfi is the state of femtocells existing in eNB i}
end if
Noting that Sf (t) is the state of the femtocell at an instant t given by:

1 if ON
Sf (t) = (10)
0 if OF F
• Step 3: After taking the decision of Switch ON/OFF, we should specify new
assignment of UEs of switched off eNB i. The allocation of a UE j is taken
depending on its SINR value. The SINR of a UE j is calculated using the
following formula:
PRX (j)
SIN R(j) = N (11)
i=j PRX (i) + KTp W
N
Where, PRX (j) is the power received at UE j, given in Eq. 8, i=j PRX (i) is
the interference average power. KTp W indicate the background noise, where
K is the Boltzmann constant, Tp is the temperature and W is the bandwidth
used by UE.
The decision to allocate each UE to the correct cell is done referring to
the Table 1.
Table 1. SINR value according to UE position in LTE-A
UE position SINR (dB)

Around the eNB ≥13
Middle of cell From 0 to 13
End of cell ≤0
The best femtocell f is chosen according to the following hypothesis:
f = arg
(f )
max (PRX (i, j)) (12)
f ∈Nf emtocells
Where, Nf emtocells (i) is the number of femtocells existing in the cell range
of eNB i and PRX is the received power at user side j coming from the
femtocell f .
The right eNB neighbor n is chosen according to the following hypothesis:
(n)
= arg
n max (PRX (i, j)) (13)
n∈Nneighbors
Where, Nneighbors (i) is the number of eNB i’s neighbors and PRX is the power
received at user side j coming from eNB neighbor n.
When we activate femtocells and attribute to them UEs of deactivated
eNB, a new amount of power denoted δPf emto is added to the total power
Ptotal . δPf emto is given by:
Nf emto

δPf emto (t) = Sf (t).Pf (14)
f =1
Pf is the power of one femtocell and Sf (t) indicate its state defined in Eq. 10.
The assignment of UEs is described in more details in the following algorithm:
Algorithm 2. Assignment of UEs

if Ni (t) ≤ T then
Si (t) ← 1
if SIN R(j) > 0 then
Sfi (t) ← 1
f ← U E(j) {UE j is attributed to femtocell f}
else if SIN R(j) ≤ 0 then
n
← U E(j) {UE j is attributed to neighbor eNB n }
{Power adjustment and tilt variation procedure}
end if
end if
• Step 4: After allocation of UEs to femtocells, UEs in the cell borders are
attributed to their nearest neighbor n . Two mechanisms are performed to
ensure coverage: a power adjustment procedure is applied in parallel with
a tilt variation of the antenna tilt of eNB neighbor n0 . The antenna tilt
variation’s process is performed in order to achieve coverage continuity and
raise cells coverages.
The better value of θ should be chosen to offer coverage for UEs located
in the cell extremity of deactivated eNB. Thus the choice of θ should respect
the following condition:
PRX (θ) < Pmin (15)
If the previous condition is false, no tilt variation and no power addition is
performed.
When we apply the tilt variation procedure, an amount of power δPtilt is

added. δPtilt depends on distance between the deactivated eNB and its neigh-
bors:
δPtilt (t) = max |Pmin − PRX (θ)| (16)
• Step 5: After the procedure of Switch ON/OFF and power adjustment and
tilt variation, the next step is to calculate the total power at instant t. ΔP
in Eq. 7 is calculated as:
ΔP (t) = δPf emto (t) + δPtilt (t) (17)
Then, the total power Ptotal (t) at time t is given by:
N Nf emto M

Ptotal (t) = [POP (i)+PT X (i).Si (t)]+ Sf (t).Pf + [Pmin −PRX (θ, k)]
i f k=1
(18)
5 Perfermance Evaluation
5.1 Simulation Scenario
To assess performances of our proposed scheme, we use an LTE-Advanced-based
HetNet topology using MATLAB as implementation software. We propose an
urban scenario where our topology is composed by N = 7 macrocells. The dis-
tance between eNBs is D = 500 m, and the eNBs-UEs distance is computed
through their positions in the map. As shown in Fig. 4, in each macro cell we
disposed two femtocells. At instant t = 0, all femtocells are in OFF state.
Fig. 4. Network architecture
For the allocation of UEs between cells, we suppose a non-uniform and ran-
dom distribution of UEs between cells where the maximum number of UEs served
by an eNB is fixed to 100 UEs. The variation of the total number of active users
during time is presented in Fig. 5.
For the simulation, we consider the eNB’s maximum transmit power for the
downlink transmission that is recommended by [14]:
Fig. 5. Number of active UEs variation Vs time
• 43 dBm for a bandwidth of 1.25, 5 MHz.

• 46/49 dBm for a bandwidth of 10, 20 MHz.
The simulation parameters employed in our simulation scenario are shown in
Table 2.
Table 2. Simulation parameters
Parameters Values
Bandwidth [MHz] 20
Frequency [MHz] 2140
Propagation model Pathloss model
Distribution of user non uniform
Transmitted power for macro cell [dBm] 35–46
Transmitted power for femto cell [dBm] 5
Distance between macro eNBs [m] 500
Number of resource blocks (RBs) in macro cell 100
Number of resource blocks (RBs) in femto cell 20

In this section, we will present simulation results and evaluation of the proposed
scheme.
Starting by Fig. 6, representing the behavior of energy consumption in the
network over time for different Switch ON/OFF algorithms. In this figure, we
compare our proposed algorithm by several cases: the first case represented by
the plot in blue describe the energy consumption without the use of any energy
saving algorithm. The second curve in green represents the case of random Switch
ON/OFF [6] where we chose to turn off 1/3 of BSs randomly. The curve in red
describes the case of Switch ON/OFF at low traffic [9]. Finally, the pink plot
Fig. 6. Power consumption Vs time
represents the energy consumption after the use of our proposed Switch ON/OFF
algorithm.
Looking at the Fig. 6, we notice that in high traffic hours, our proposed
algorithm gives almost the same results as the low traffic algorithm, but in
periods of low traffic, our proposed approach realizes the better power saving
than the case of no switch ON/OFF technique is performed and the cases of low
traffic and random schemes.
The percentage of power saving is the result of Fig. 7. As shown in this figure,
the proposed scheme achieves the maximum amount of power saving than the
random and the low traffic schemes. The percentage of power saving achieved by
our proposed scheme can attain 32% especially during night period from 00:00
am to 05:00 am.
Fig. 7. Power saving (%) Vs time
The use of femtocells improves the amount of energy saving; thus, femtocells
are characterized by their low amount of energy consumption. Better yet, the use
of the tilt variation technique increase the amount of power saving after applying
the switch ON/OFF algorithm, thus, it improves the QoS by guarantying an
acceptable coverage to UEs.
In order to evaluate QoS and coverage, we simulate SINR values before and
after the Switch ON/OFF procedure. Using Eq. 11, we calculate SINR for UEs
existing in the cell edge. The result of SINR calculation is presented in Fig. 8. As
it is shown, the SINR values after the switch ON/OFF procedure increase and
the coverage is guaranteed, Using the concept of femtocells and eNB parameters
tuning let the QoS stable and not lot affected.
Fig. 8. SINR values before and after the Switch ON/OFF procedure
6 Conclusion
The purpose of our work in this paper is to find a better method that can reduce
energy consumption in a LTE-A based HetNet network. Our algorithm is based
on two concept: the first one is the use of femtocells that can help to reduce
the energy consumption in the network thanks to their low energy consumption.
The second concept is the tilt variation technique that ensures an acceptable
coverage and maintains QoS. According to the SINR values before and after
Switch ON/OFF, we can say that our algorithm does not affect the QoS of the
network.
Acknowledgment. This work has received funding from the European Union’s Hori-
zon 2020 research and innovation programme under grant agreement No. 761579 (TER-
APOD).
References
1. http://www.3gpp.org/hetnet. Accessed 15 Nov 2018
2. Global Action Plan Report, An inefficient truth, December 2017
3. http://green-broadband.blogspot.com. Accessed 15 Nov 2018
4. Han, C., et al.: Green radio: radio techniques to enable energy efficient wireless
networks. IEEE Commun. Mag. 49(6), 46–54 (2011)
5. Gong, J., Zhou, S., Niu, Z., Yang, P.: Traffic-aware base station sleeping in dense
cellular networks. In: IWQoS 2018, pp. 1–2, June 2010
6. Chiaraviglio, L., Ciullo, D., Meo, M., Marsan, M.A.: Energy-aware UMTS access
networks. In: W-GREEN 2008, Lapland (2008)
7. Chiaraviglio, L., Ciullo, D., Meo, M., Marsan, M.A.: Energy efficient management
of UMTS access networks. In: ITC 21, Paris, France, pp. 1–8, September 2009
8. Bousia, A., Kartsakli, E., Alonso, L., Verikoukis, C.: Energy efficient base station
maximization switch off scheme for LTE-advanced. In: IEEE International Work-
shop CAMAD, Barcelona, Spain, pp. 256–260, September 2012
9. Lassoued, N., Boujnah, N., Bouallegue, R.: Power saving approach in LTE using
switching ON/OFF eNodeB and Power UP/DOWN of Neighbors, SETIT, Tunisia,
December 2018
10. Lassoued, N., Boujnah, N., Bouallegue, R.: Reducing power consumption in HetNet
network using power adjustment and coordinated multipoint technique. In: AINA,
Cracow, Poland, pp. 539–544, May 2018
11. Andrews, J.G., Claussen, H., Dohler, M., Rangan, S., Reed, M.C.: Femtocells: past,
present, and future. IEEE J. Sel. Areas Commun. 30(3), 497–508 (2012)
12. Guidelines for evaluation of radio interface technologies for IMT-Advanced. Radio-
communication Sector of ITU (2009)
13. Salous, S.: Radio Propagation Measurement and Channel Modeling. Wiley,
New York (2013)
14. E-UTRA; further advancements for E-UTRA physical layer aspects, 3GPP (2010)
Robust Road Lane Detection for High
Speed Driving of Autonomous Vehicles
Hyunhee Park(B)
Department of Computer Software, Korean Bible University, Seoul, South Korea

parkhyunhee@gmail.com
Abstract. With the Hough transform and region of interest, how to

improve the processing time of high-speed driving is being actively inves-
tigated. This study proposes a road lane detection algorithm based on
expressway driving videos through a computer vision-based image pro-
cessing system without using sensors. The proposed method detects
straight lines that are estimated to be lanes using the Hough trans-
form. When lanes are detected from actual images, the scope of left and
right lanes is limited to reduce computational load. Extensive simula-
tion results are given to show the effects of Hough transform method for
high speed driving and region of interest for processing time on actual
expressways.
1 Introduction
Various technologies are being researched to improve the safety and convenience
of vehicle driving [1]. In particular, technologies using image information include
pedestrian detection, vehicle detection, obstacle detection, and lane detection
[2,3]. Among these, lane detection techniques have become a core technology
for autonomous vehicles as they can be directly applied to lane departure pre-
vention and lane keeping technologies [4]. Actual lanes are mixtures of straight
and curved lanes which need to be detected quickly. Image-based lane detection
technologies use different algorithms for straight and curved lanes. Therefore,
techniques to quickly predict straight and curved lanes and to determine lanes
are required. The existing methods using straight line models are mainly used
to detect nearby lanes that appear almost straight. Furthermore, methods that
consider curved sections as successions of straight sections divide a lane into
multiple sections by distance; the lanes in the curved sections are then detected
using the fact that each short section would be approximately straight [5,6].
When lanes are assumed to be straight sections, the Hough transform model is
typically used to detect straight lanes. Applying the Hough transform enables
the effective detection of various straight lines in images containing lanes. The
Hough transform method converts points in an image space into a Hough space,
and chooses a section that is highly likely to be a straight line by comparing the
accumulated values. The conventional Hough transform method converts all fea-
ture points in the entire image into a Hough space. When the Hough transform
https://doi.org/10.1007/978-3-030-15035-8_24
Robust Road Lane Detection for High Speed Driving 257
is performed in this way, the Hough space creation step takes too much time,
and unnecessary data can be detected. Therefore, to detect lanes quickly during
high speed driving, this study proposes a lane detection method that applies the
vanishing point (VP) update method considering the camera location relative
to the road surfaces and lanes in the Hough space creation process [7,8]. The
proposed method updates the VPs by accumulating only information related to
straight lines passing near the VPs and performs the Hough transform only for
straight lines that pass around the predicted VPs. A model that is appropriate
for high speed driving is then created by predicting the areas in advance for
which the Hough transform will be carried out without measuring all straight
lines and distances.
2 Robust Hough Transform Method for High Speed

Driving
Figure 1 shows the road image used for lane detection. In general, the horizontal
lines and VPs are located at the center or a little above the center of the image
[9]. A VP here means a point at which the parallel straight lines on the same
plane meet after a projection transformation in the camera. If lane changes in a
general road driving environment are excluded, VPs are located within a certain
area in the images. However, the VPs in a curved section meet at different
points. If a curved line is regarded as a collection of sufficiently short straight
lines, when a lane is divided horizontally into a few sections and then they are
smoothly interconnected, a curved line is created.
Figure 2 shows the first VP result created in a curved section. From the VP
on the triangle formed by the left bottom point, the VP, and the right bottom
point, the height of the triangle was divided by half, and the lane was detected
again. The result is shown in Fig. 3, which is the result of the second VP created
in a curved section. Using the same process, the third VP was created, which is
shown in Fig. 4. When these VPs are processed into Bezier curves and smoothly
interconnected, the VPs for a curve can be formed naturally as the VPs move.
One thing that needs to be considered here is the region of interest (ROI).
For fast image processing and VP prediction, the ROI area is made to decrease
step by step, thereby substantially reducing the computational load. Figure 5
shows the ROI extraction result based on the VPs for the entire image. The
computational load of the operation can be significantly reduced by designating
the red area as the ROI based on the VPs. Figure 6 shows the results of edge
detection by ROI. As a preprocessing step, the noise generated during the edge
extraction was removed with a Gaussian filter (5 * 5) and the image was further
smoothed using a median filter. Then, four points were created by selecting four
coordinates using the ROI, and a polygon was created based on these points. The
computational load was reduced and the accuracy was improved by allowing only
the image inside this polygon to be detected. The processing required for edge
detection can be significantly decreased by reducing the operations used for the
surrounding environment, including lanes in the vicinity. Furthermore, as shown
258 H. Park
Fig. 1. Example of a road image with VP.
Fig. 2. Example of the first VP result created in a curved section.
in Fig. 7, the second ROI is determined based on the second VP creation area in
which the VP moves. This is reduced to half the size of the first ROI. In other
words, the computational load can be halved compared to the first ROI. Through
the same process, the third ROI in Fig. 8 is determined. Finally, a straight line
is obtained using the Hough transform and the lane is detected and displayed
on the screen.
Fig. 3. Example of the second VP result created in a curved section.
Fig. 4. Example of the third VP result created in a curved section.
3 Using Continuous Frames to High Speed Driving

To determine a lane using the Hough transform from road images that are input
continuously, the following method is used. In this study, before the lane is
determined, the left and right lanes are determined first based on the driving
data. After the Hough transform is performed based on the ROI and VPs, the
point with the largest accumulated value in the area [0, π/2) is determined as
the left lane, and the largest point in the area (π/2, π] is determined as the
right lane. When (1) is applied in a continuous image input situation, the VP
260 H. Park
Fig. 5. Example of ROI extraction result based on the VPs.
Fig. 6. Result of edge detection by ROI.
acquired from the previous image is used. In this case, the center of the first
image is assumed to be the VP of the previous image. The following equation
used for Hough space generation for the nth image:
γ = (x − xn−1
v ) cos θ + (y − yvn−1 ) sin θ. (1)
Fig. 7. Example of second ROI extraction based on the second VP.
Fig. 8. Example of third ROI extraction result based on the third VP.
To measure the performance of the proposed Hough transform method, the fol-
lowing experiment was performed. The main goal of the proposed method is to
improve the lane detection speed using Hough transform. Therefore, the Hough
transform speed and lane detection speed were measured and compared with
those of existing methods. Furthermore, the lane detection performance is com-
pared with the existing methods to measure the change in accuracy according
to the speed improvement. All the images used in this experiment were taken
262 H. Park
Fig. 9. Hough space processing time by noisy point (ms).
from actual roads and the image size was 960 * 540 pixels. To detect lane fea-
tures from actual road images, a total of 6000 images were obtained including
2000 images each for straight, curved, and lane change sections. For objective
performance measurement, randomly created virtual images were also used and
compared. These images correspond to the preprocessed results of actual road
driving images. Virtual images were used for accuracy measurement and com-
parison. The proportions of straight lines with a distance of 50 pixels or less
from the VP to the straight line were set to 10%, 50%, and 90%. For the slopes
of straight lines, 22 slopes from −4 to 4 were used so that all the components
of the coordinates would be integers. Among these, 10 slopes were randomly
selected and straight lines were drawn. In addition, a total of 6000 images were
created by marking 0, 500, 1000, 1500, and 2000 points at random positions
in each image, similar to the actual situation. Because fewer feature points are
detected in remote distances, zero to 2000 noise elements were added to the
images. In other words, 2000 noise elements were generated for each straight
line. The reason for adding noise is that the results obtained after preprocessing
from the road images are not only lane features. In other words, the amount of
noise can vary by the acquired image and the performance of the preprocessing
module. Therefore, an experimental set of various noise intensities was created
and used to demonstrate that the proposed method is also robust against the
effect of noise regardless of external factors. In Fig. 9, even if the amount of noise
is increased, the processing time of the proposed method increases slowly, while
it increases rapidly in other methods.
The performance of each method was compared based on the Hough space
generation speed measured using actual road images. The average measurement
time was determined using 6000 images for each method. The experimental
results in Fig. 10 show that the proposed method is robust against noise in the
Fig. 10. Hough space processing time by various road environment (ms).
Fig. 11. Lane detection time in previous mechanism (ms).
Hough space generation process and faster than the other methods. The experi-
mental results for the Hough space generation and lane detection process using
6000 images are as follows. The processing time of each method was measured
for 1000 road images of each type, and the average processing times are shown
in Figs. 11 and 12. The computation speed of the proposed method was faster by
at least 2 times and to up to 20 times compared to other methods in all cases.
This result suggests that the proposed method is appropriate for detecting lanes
in road images under high speed driving conditions.
264 H. Park
Fig. 12. Lane detection time in proposed mechanism (ms).
5 Conclusion
This study proposed a modified Hough transform method and a lane detection
method in a Hough space. The proposed method only generates a Hough space
with straight lines adjacent to the VP of the previous frame. The size of the
Hough space can be reduced because it is generated based on the ROI, thus
significantly reducing the computation time for the Hough transform and lane
detection. The experimental results show that the computation speed of the
proposed method in the Hough transform and lane detection process is approx-
imately twice that of the existing methods. The method proposed in this study
can be useful for lane detection systems installed in vehicles due to its high
computation speed and low memory use.
Acknowledgements. This work was supported by the National Research Founda-

tion of Korea (NRF) grant funded by the Korea government (MSIT) (No. NRF-
2017R1C1B5017556).
References
1. Park, H.: Implementation of lane detection for self-driving vehicles using GPS. J.
Korean Inst. Commun. Inf. Sci. 43, 18 (2018)
2. Labayrade, R., Douret, J., Aubert, D.: A multi-model lane detector that handles
road singularities. In: Proceedings of the IEEE Conference on Intelligent Trans-
portation Systems, Toronto, Canada, pp. 1143–1148 (2006)
3. McDonald, J., Shorten, R., Franz, J.: Application of the Hough transform to lane
detection in motorway driving scenarios. In: Proceedings of the Irish Signals and
Systems Conference (2001)
4. Yadav, S., Patra, S., Arora, C., Banerjee, S.: Deep CNN with color lines model for
unmarked road segmentation. In: Proceedings of the IEEE International Conference
on Image Processing (2017)
5. Guo, C., Mita, S., McAllester, D.: Robust road detection and tracking in challenging
scenarios based on Markov random fields with unsupervised learning. IEEE Trans.
ITS 13, 1338–1354 (2012)
6. Tsai, L., Hsieh, J., Chuang, C., Fan, K.: Lane detection using directional random
walks. In: Proceedings of IEEE Intelligent Vehicles Symposium (2008)
7. Li, Q., Zheng, N., Cheng, H.: Spring robot: a prototype autonomous vehicle and its
algorithms for lane detection. IEEE Trans. Intell. Transp. Syst. 5, 300–308 (2004)
8. Wang, J., Ji, Z., Su, Y.: Unstructured road detection using hybrid features. In:
Proceedings of the International Conference on Machine Learning and Cybernetics
(2009)
9. Yun, S., Guo-Ying, Z., Yong, Y.: A road detection algorithm by boosting using
feature combination. In: Proceedings of the Intelligent Vehicles Symposium (2007)
Study of Beam Power Control
of Ka-Band Multi-beam Broadcasting
Satellite Using Meteorological Data
Takumi Iwamoto and Kiyotaka Fujisaki(B)
Fukuoka Institute of Technology, 3-30-1 Wajiro-higashi,

Higashi-ku, Fukuoka 811–0295, Japan
mgm18101@bene.fit.ac.jp, fujisaki@fit.ac.jp
Abstract. Recently, higher definition broadcasting services such as

4K/8K is required in the satellite link. High-speed and large-capacity
satellite link will be required to achieve this demand in the future. One of
the means to achieve such link is the use of higher frequency band such as
Ka-band (30/20 GHz) for the next-generation broadcasting. However, the
attenuation caused by rain gets large in accordance with the frequency.
Therefore, it is essential to develop the rain fade mitigation technique
as a countermeasure against the rain attenuation. For this reasons, we
suggest an adaptive control method using Ka-band multi-beam satellite.
In the proposed method, the satellite link is controlled adaptively based
on the weather condition. We focus on the Rader-AMeDAS precipitation
data as meteorological data for the beam control.
1 Introduction
Japanese satellite broadcasting services of the standard definition resolution

using Ku-band (17/12 GHz) started in 1989. Nowadays the services of high defini-
tion resolution are offered. On the other hand, for the next-generation of higher
definition broadcasting, 21 GHz (Ka-band) is assigned to Asia region in 1992.
After the assignment of Ka-band, broadcasting services continues to be devel-
oped, and Japanese broadcasting company has started 4K/8K services using
current Ku-band in 2018. The Ku-band has been used in the current satellite
broadcasting until now, but this frequency band cannot accommodate any more
services. Despite this fact, 4K/8K services has been achieved with the Ku-band
by sharing right and left-handed circular polarization, reallocating the assigned
frequency band and so on. One of the major reasons why the next-generation of
Ka-band is not used is because still there is not any effective rain fading mit-
igation technique. In the previous research, the conventional countermeasures
against the rain attenuation give satellite link large electricity as the rain mar-
gin. However, it is impossible for the next-generation of the Ka-band satellite to
give the sufficient rain margin because of following reasons:

https://doi.org/10.1007/978-3-030-15035-8_25
Study of Beam Power Control of Ka-Band Multi-beam Satellite 267
• Dramatical increase of rain attenuation in comparison with Ku-band,

• Strict limitation of satellite electricity.
Even if we give sufficient rain margin, the most electricity will be wasted in case
of fine weather.
Therefore, we focus on an adaptive control method using multi-beam satellite
[1–3]. In our proposed method, first, the rain effect of the next time on the
satellite link is predicted based on the weather condition in each area. Next,
the transmission power of several beams is controlled based on the prediction.
We have considered the effectiveness of the beam control method under the ideal
condition [4,5]. In the proposed method, it is important to consider the precision
of precipitation data. Therefore, in this paper, we consider the precipitation data
of Rader-AMeDAS (Automated Meteorological Data Acquisition System) for the
prediction.
The structure of this paper is as follows. In Sect. 2, we give the outline of the
proposed method and the satellite model. Based on the explanation, we show the
detail simulation procedure, the method to predict link quality and method to
evaluate the link quality. In Sect. 3, the simulation result in case of using Rader-
AMeDAS data has been shown to consider the effectiveness of the precipitation
data. In Sect. 4, we conclude this paper.
(a) Single-beam satellite (b) Multi-beam satellite

(Conventional) (Next-generation)
Fig. 1. Difference between conventional satellite system and next-generation satellite

system.
2 Adaptive Control Method and Evaluation Method
We focus on the adaptive control method using the multi-beam satellite as a

countermeasure against the rain attenuation. As shown in Fig. 1, while the con-
ventional single-beam satellite covers area with a large beam, the next-generation
of multi-beam satellite does with several beam, and we can control the commu-
nication parameter of each beam. For example, it becomes possible to change
268 T. Iwamoto and K. Fujisaki
the transmission power of each beam [6]. Therefore, it can be considered that
effective countermeasure against the rain attenuation would be achieved by
changing the transmission power of each beam based on the weather condition
in each area.
Beam 7
Beam 8
Beam 0
Beam 6
Beam 5
Beam 3
Beam 2
Beam 4
Beam 1
Fig. 2. The beam placement in the simulation.
2.1 Simulation Procedure
We model the simulation based on WINDS (Wideband InterNetworking engi-

neering test and Demonstration satellite) launched in 2007. WINDS has been
developed in Japan to demonstrate the techniques of ultra-high-speed Inter-
net by using Ka-band and multi-beam technique. In our simulation, nine multi
beams are placed so that they cover all area of Japan in reference to the WINDS
as shown in Fig. 2 [7]. The simulation flow is as follows.
1. The precipitation data for the prediction of the link quality of next time
obtained by Rader-AMeDAS is gathered.
2. Based on the data, several beams whose link quality is degraded due to
the rain are selected and the transmission power toward of those beams is
enhanced in addition to the normal power.
3. For the evaluation of the link quality, Rader-AMeDAS measured precipitation
data are used.
(a) Japan (b) South Japan
Fig. 3. Example of Rader-AMeDAS precipitation image data.
2.2 Prediction of Link Quality Method
In order to predict the rain effect on the link quality, we use following steps.
1. Beam selection method

By using Eq. (1), the rain effect on the link in each beam is evaluated, and the
beam to enhance the electricity is selected based on the order of the average
rainfall.
Total amount rainfall of 15 − 25 mm/h

Average rainfall = (1)
Total number of observation point
Here, the precipitation of 15–25 mm/h can be relieved by enhancing the elec-
tricity, and we call the selected beam “enhanced beam” in this simulation.
2. Precipitation data to control the beams
Figure 3 is the example of Rader-AMeDAS precipitation image data used in
the simulation and the difference of the color shows the difference of precip-
itation. In the simulation, we take advantage of the predicted precipitation
image data or that of the measured image data from Rader-AMeDAS to pre-
dict the link quality of the next time as shown in Fig. 4. The ideal case in
Fig. 4 is that the precipitation data used for the prediction of the link quality
and actual measurement are the same. Therefore, the link quality under this
condition can be regarded as the target link quality.
2.3 Evaluation of Link Quality Method
In this research, we consider the following condition for the evaluation of the
link quality.
Current Time
10:00 11:00 12:00
Example 1)
In case of using past 1 hour Measured data
measured precipitation data from 10:00 to 11:00
Example 2)
Predicted data Prediction of link
In case of using after 1 hour
from 11:00 to 12:00 quality of the next time
predicted precipitation data
Example3)
Ideal case of using the time Measured data
precipitation data (ideal case) from 11:00 to 12:00
Fig. 4. The use of precipitation data.
1. Threshold rainfall
According to the ITU-R recommendations [8], the specific rain attenuation
γ [dB/km] is obtained from;
γ = kRα (2)
Here, R [mm/h] is rain rate. k and α are frequency-dependent coefficients.
In order to evaluate the link quality, we define the threshold rainfall. If the
precipitation in a observation point exceeds the threshold rainfall, the point
is regarded as the incommunicable state. The threshold rainfall shown in
Table 2 is calculated by using Eq. (2) and the parameters shown in Table 1.
The threshold rainfall in case of “Not-enhanced beam” is 7.5 mm/h. On the
other hand, because the electricity is enhanced in addition to the normal
power, the threshold rainfall is 25 mm/h.
2. Beam settings
The diameter of the beam is set to 450 km. The precipitation image data used
for controlling the beam consists of 640 × 640 pixels, and each beam shown in
Fig. 2 has about 20,000 pixels. In the simulation, the pixel in the image data
is assumed as the observation point of the radio wave from the satellite.
3. Evaluation of the link quality
Figure 5 shows the simulation flow for the evaluation of the link quality. In
each observation point (=pixel), based on Eq. (3), the link quality is evaluated
as follows;

rainfall > threshold rainfall : Incommunicable state
(3)
rainfall ≤ threshold rainfall : Communicable state
The threshold rainfall is shown in Table 2 and the availability is defined as a

link quality as follows;
The number of communicable observation point
Availability[%] = × 100.
The number of total observation point
(4)
Table 1. Simulation specifications.
Launch orbit Geostationary orbit

Launch position 143◦ east longitude
Frequency 20 GHz
k (at 20 GHz) 0.0814
α (at 20 GHz) 1.0754
0◦ height 4.3 km
Modulation schemes TC8PSK
Rain margin 6.3 dB
Table 2. Threshold rainfall.
Power distribution Threshold rainfall

Not-enhanced beam 11.1 W (= 100 W/9) 7.5 mm/h
enhanced beam 44.4 W (= 100 W/9 + 100 W/)4 25 mm/h
The precipitation image data AMeDAS precipitation

of each observation point is image data is used.
obtained.
Average rainfall selects
Several beams whose link beams.
quality degrades due to the
rain are selected. The number of enhance
beam is 3.
Enhance beam is selected
The transmit power of those
at 1 hour intervals.
beams is enhanced in addition
to the normal power. As a link quality, the link
availability is defined, and
The link availability is that of each beam is
evaluated. evaluated at 1hour intervals.
Fig. 5. Simulation flow.

Precipitation data obtained in 2017 is used for the simulation. Measured precip-
itation data or predicted precipitation data at most 2 h before current time is
used for predicting the link quality.
3.1 Annual Availability
The availability obtained by using the adaptive control of beams with the pro-
posed technique is shown in Fig. 6. The availability becomes 99.55% in case of
no-countermeasure. From this figure, it has been shown that:
• the annual availability is improved compared with no-countermeasure,

• the availability degrades as the old data is used for the prediction of the link
quality,
• in case of using data observed 1 h ago, the availability degrades only about
0.01% compared with ideal one,
• the difference between the measured data and predicted data becomes big for
old data.
3.2 Beam Availability
First, we evaluate the annual availability of each beam in case of using measured
precipitation data. The results are shown in Fig. 7. From the results, it has been
shown that:
• the availability of the beams on the south such as beam 0 and beam 1, covering
southern area where it has much rain, gets greatly improved compared with
the no-countermeasure,
• similar to the case of annual availability, the availability degrades as the
precipitation data becomes old.
Next, we evaluate the annual availability of each beam in case of using pre-
dicted precipitation data. The results are shown in Fig. 8. From the results, it
has been shown that:
• the availability tendency is same as case of measured data, though there exists
some variations,
• the degree of degradation of availability in case of using predicted data is
smaller than that of using measured data.
Measured data Predicted data

99.93
99.91469 99.91721
99.92 99.92612
99.90865
Availability[%]
99.91 99.90184
99.90
99.89
99.88
99.87
99.86
ideal 1 hours ago 2 hours ago
Fig. 6. Annual availability (2017).
ideal 1 hour ago 2 hours ago No-countermeasure

100.00
99.80
Availability[%]
99.60
99.40
99.20
99.00
98.80
0 1 2 3 4 5 6 7 8
Beam number
Fig. 7. Beam annual availability (Measured data, 2017).
ideal 1 hour ago 2 hours ago No-countermeasure

100.00
99.80
Availability[%]
99.60
99.40
99.20
99.00
98.80
0 1 2 3 4 5 6 7 8
Beam number
Fig. 8. Beam annual availability (Predicted data, 2017).

4 Conclusions
In this paper, as a countermeasure against the rain attenuation of the next-
generation broadcasting satellite system using Ka-band, we suggest the technique
using multi-beam satellite. In the proposed method, the multi-beam is controlled
based on the weather condition in each area. We considered the AMeDAS pre-
cipitation image data measured in the past or predicted in the past for the beam
control.
The simulation results show that the availability degrades when using old
data. However, the availability degrades only about 0.1% in the case of using the
past 1 h data. Therefore, it seems that we can control the beams appropriately
by the past 1 h data.
For the future work, we will consider 10 min intervals, such as AMeDAS
10 min precipitation data, in order to realize a situation more similar to the
reality.
References
1. Yoshino, T., Ito, S.: A selection method of beams compensated using AMeDAS data
for the multi-beam satellite broadcasting. IEICE Trans. B J82-B(1), 64–70 (1999).
(in Japanese)
2. Matsudo, T., Karasawa, Y., Shiokawa, T.: The effect of compensation for rain atten-
uation on Earth-satellite links by onboard resource sharing using AMeDAS data.
IEICE Trans. B J76-B2(5), 373–381 (1993). (in Japanese)
3. Fukuchi, H., Sugimoto, T., Tsuzuku, A.: Advanced satellite broadcasting system in
the 21GHz band -Quantitative evaluation of the effect of adaptive satellite power
control. ITE Tech. Rep. 18(45), 13–19 (1994). (in Japanese)
4. Fujisaki, K.: Study on Improvement of Link Quality using AMeDAS precipitation
data for Multibeam Satellite System. “Evaluation of the Down Link Availability
under the Most Suitable Beam Control”. Fukuoka Inst. Technol. Bull. Paper 28,
9–14 (2017). (in Japanese)
5. Iwamoto, T., Fujisaki, K.: Study of the improvement of the link quality using
AMeDAS rainfall data for multibeam satellite broadcasting system. IEICE Techn.
Rep. 118(176), 85–90 (2018). (in Japanese)
6. Kadowaki, N., Suzuki, R.: Overview of the wideband internet engineering test and
demonstration satellite project. J. NICT 54(4), 3–10 (2007)
7. Ozawa, S., Hasegawa, T., Shimada, M., Hirayama, K., Koishi, Y.: Multibeam
antenna system. J. NICT 54(4), 35–43 (2007)
8. Rec. ITU-R P.618-8, Propagation data and prediction methods required for the
design of Earth-space telecommunication systems, ITU (2001)
Numerical Analysis of Optical Duplexer
Composed of Dispersive and Nonlinear
Dielectric in Two-Dimensional Photonic
Crystal Waveguide with Square Lattice
Keisuke Haari1(&), Naoki Higashinaka1, Xiang Zheng Meng1,

and Hiroshi Maeda2
1
Graduate School of Communication and Information Networking,
Fukuoka Institute of Technology, Fukuoka, Japan
mgm17105@bene.fit.ac.jp
2
Department of Information and Communication Engineering, Fukuoka, Japan
hiroshi@fit.ac.jp
Abstract. For optical wavelength division multiplexing (WDM) signal, optical

duplexer composed of square-lattice photonic crystal structure with nonlinear
and dispersive material was analyzed numerically. The output power distribu-
tion in two output ports was compared with duplexer by linear dielectric
material. It was found that output spectrum changes in case of nonlinear Kerr
medium compared with linear case. The result suggests that optical signal
switching is available depending on signal intensity. This is useful for all-optical
signal switching for optical WDM telecommunication system.
1 Introduction
Photonic crystal (PC) structure is very important for the integration of optical circuit
because of its propagation characteristics based on the photonic band gap (PBG). The
pillar type PC waveguide is basically composed of periodic array of dielectric rods with
a line of defect in free space, as illustrated in Fig. 1. Because of its band gap, the wave
with limited propagation constant can exists in the waveguide. Its propagation char-
acteristics are determined by the refractive index, the lattice period, and optical fre-
quency [1, 2].
Nonlinear optical effect shows many interesting behaviors. Especially, optical Kerr
effect shows self-modulation of the refractive index and self-focusing/defocusing. The
propagation constant is changed by its optical field.
In two dimensional PC structure, the dispersive linear and nonlinear optical
duplexer is composed of waveguide. Use silica of dispersive linear and nonlinear
dielectric material, and consider whether route-switching is possible depending on the
wavelength of input optical signal.
To investigate such PC waveguide, the finite difference time domain (FDTD)
method is used. The FDTD method can be applied to dispersive linear and nonlinear
dielectric material [3].

https://doi.org/10.1007/978-3-030-15035-8_26
276 K. Haari et al.
Fig. 1. Illustration of optical duplexer with photonic crystal waveguide.
In this paper, analysis of the optical switching by FDTD method in the PC

waveguide with optical duplexer is reported.
2 Finite Difference Time Domain Method

@
According to Yee’s FDTD scheme [4], the two dimensional (i.e. @y ¼ 0), discretized
Maxwell’s equation for transverse electric (TE) mode, which propagates to axis, are
described as follows;
ð1Þ
ð2Þ
ð3Þ
where E and H are electric and magnetic field and D is the electric displacement. The
space and time discretization are and i; k; n are the number of corresponding
grid. Giving a set of suitable initial conditions and a constitutive equation between
Numerical Analysis of Optical Duplexer Composed of Dispersive 277
D and E to above equations, the latest field is calculated successively as the increase of
time step number.
Consider composite space of free space, linear dielectric material, and dispersive
linear and nonlinear dielectric material.
The electric field in free space is described as follows;
Dny
Eyn ¼ ð4Þ
e0
where e0 is permittivity of free space.

The electric field in linear dielectric material is described as follows;
Dny
Eyn ¼ ð5Þ
e0 er
where er is relative permittivity of linear dielectric material.

The electric field in dispersive linear and nonlinear dielectric material is described
as follows [3];
3
ð3Þ
Dny Sn1
1
e0 L þ 2v0 a Eyn1
Eyn ¼ 2 ð6Þ
ð3Þ ð3Þ
e1 þ v0 ð1 aÞSn1
R þ 3v0 a Ey
n1
SnL ¼ 2 eaL Dt cosðbL DtÞ Sn1

L e2aL Dt Sn2
L
ð7Þ
þ cL Dt eaL Dt sinðbL DtÞ En
SnR ¼ 2 eaR Dt cosðbB DtÞ Sn1

R e2aR Dt Sn2
R
ð8Þ
þ cR Dt eaR Dt sinðbR DtÞ ðEn Þ2
aL ¼ 2pfL dL ð9Þ
qffiffiffiffiffiffiffiffiffiffiffiffiffi
bL ¼ 2pfL 1 d2L ð10Þ
2pfL ðes e1 Þ
cL ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffi ð11Þ
1 d2L
aR ¼ 2pfNL dNL ð12Þ

qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
bR ¼ 2pfNL 1 d2NL ð13Þ
2pfNL vð3Þ ð1 aÞ
cR ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ð14Þ
1 d2NL
278 K. Haari et al.
where e1 and es are relative permittivity of dispersive linear and nonlinear dielectric
material, fL is linear relaxation frequency, dL is linear decaying factor, fNL is nonlinear
ð3Þ
relaxation frequency, dNL is nonlinear decaying factor, v0 is nonlinear susceptibility
and a is weight factor for Kerr effect.
3 Numerical Examples
The distribution ratio of each PORT, the extinction ratio of the dispersive linear and
nonlinear optical duplexer, and the electric field distribution of the photonic crystal
waveguide were obtained. PORT 1 is a PORT to confirm reflection, PORT 2 and
PORT 3 are output PORT. Three types of dispersive linear and nonlinear optical
duplexer structures were used.
The dispersive linear and nonlinear optical duplexer structure is shown in Fig. 1.
Black circle is the silicon of linear dielectric material, blue circle is the silica of
dispersive linear and nonlinear dielectric material. The refractive index of linear
dielectric material n1 ¼ 3:6 as is in Ref. [5]. The refractive index of dispersive linear
and nonlinear dielectric material n2 ¼ 1:5, relative permittivity of dispersive linear and
nonlinear dielectric material e1 ¼ 2:25; es ¼ 5:25, linear relaxation frequency
fL ¼ 63:7 ½THz, linear decaying factor dL ¼ 2:5 104 , nonlinear relaxation fre-
quency fNL ¼ 14:8 ½THz, nonlinear decaying factor dNL ¼ 3:36 101 , nonlinear
ð3Þ
susceptibility v0 ¼ 0:07, and weight factor for Kerr effect a ¼ 0:7 as is in Ref. [3].
The permittivity of free space e ¼ e0 , i.e., no ¼ 1:0. The lattice period a ¼ 551:8 ½nm,
the radius of dielectric material R ¼ 0:2a, the width of waveguide d ¼ 866:8 ½nm are
used. The FDTD discretizations are Dx ¼ Dz ; 9:85 ½nm and Dt ; 0:0209 ½fs,
respectively. Analysis area is 10:244 ½lm 14:1052 ½lm. As an absorbing boundary,
Berenger’s perfectly matched layer is installed.
The input optical signal is Gaussian beam and given by
( )
x 2 2pct
Ey ðx; tÞ ¼ E0 exp sin ð15Þ
w0 k
Ey ðx; tÞ
Hx ðx; tÞ ¼ qffiffiffiffi ð16Þ
e0 le00
where the amplitude E0 ¼ 1:0 ½V=m, the width of beam spot w0 ¼ 275:8 ½nm, speed
of optical in vacuum c ¼ 2:998 108 ½m=s, the wavelength k ¼ 1:350 1:600 ½lm,
the permittivity of free space e0 ¼ 8:854 1012 ½F=m, and the magnetic permittivity
of free space l0 ¼ 4p 1017 ½H=m.
The distribution ratio of each PORT is described as follows;
PPORTi
Si ¼ 100 ði ¼ 1; 2; 3Þ ð17Þ
PPORT1 þ PPORT2 þ PPORT3
where P is the optical intensity of each PORT.

The extinction ratio of the dispersive linear and nonlinear optical duplexer is
described as follows;
S2
10 log10 ð18Þ
S3
3.1 Dispersive Linear and Nonlinear Optical Duplexer 1

The first dispersive linear and nonlinear optical duplexer structure is shown in Fig. 1.
Figure 2 shows the distribution ratio, Fig. 3 shows the extinction ratio.
At the wavelength k ¼ 1:390 ½lm, the distribution ratio of PORT 2 is
S2 ¼ 86:864½%, the distribution ratio of PORT 3 is S3 ¼ 12:487½%, and the distri-
bution ratio of PORT 1 is S1 ¼ 0:649½%. The electric field distribution is shown in
Fig. 4. The extinction ratio is 8:424 ½dB.
100.000
90.000
80.000
Distribution ratio [%]
70.000
60.000
50.000
40.000
30.000
20.000
10.000
0.000
1.350E-06 1.400E-06 1.450E-06 1.500E-06 1.550E-06 1.600E-06
Wavelength [m]
PORT1 PORT2 PORT3
Fig. 2. Distribution ratio of each PORT in Fig. 1.

The wavelength difference of switching the route is Dk ¼ 163 ½nm.
280 K. Haari et al.
10.000
8.000
6.000
Extinction ratio [dB]
4.000
2.000
0.000
-2.000
-4.000
-6.000
-8.000
-10.000
1.350E-06 1.400E-06 1.450E-06 1.500E-06 1.550E-06 1.600E-06
Wavelength [m]
Fig. 3. Extinction ratio of dispersive linear and nonlinear optical duplexer in Fig. 1.
(a) (b)
Fig. 4. Electric field distribution at (a) k ¼ 1:390 ½lm and (b) k ¼ 1:553 ½lm. Coordinate axis
is in Fig. 1.

The second dispersive linear and nonlinear optical duplexer structure is shown in
Fig. 5.
Fig. 5. Dispersive linear and nonlinear optical duplexer structure.
100.000
90.000
80.000
70.000
60.000
50.000
40.000
30.000
20.000
10.000
0.000
1.350E-06 1.400E-06 1.450E-06 1.500E-06 1.550E-06 1.600E-06
Wavelength [m]
PORT1 PORT2 PORT3

282 K. Haari et al.
10.000
8.000
6.000
4.000
2.000
0.000
-2.000
-4.000
-6.000
-8.000
-10.000
1.350E-06 1.400E-06 1.450E-06 1.500E-06 1.550E-06 1.600E-06
Wavelength [m]
(a) (b)
Fig. 8. Electric field distribution at (a) k ¼ 1:515 ½lm and (b) k ¼ 1:434 ½lm.

Fig. 9. Dispersive linear and nonlinear optical duplexer structure.

100.000
90.000
80.000
70.000
60.000
50.000
40.000
30.000
20.000
10.000
0.000
1.350E-06 1.400E-06 1.450E-06 1.500E-06 1.550E-06 1.600E-06
Wavelength [m]
PORT1 PORT2 PORT3

284 K. Haari et al.

The last dispersive linear and nonlinear optical duplexer structure is shown in Fig. 9.
14.000
12.000
10.000
8.000
6.000
4.000
2.000
0.000
-2.000
-4.000
-6.000
-8.000
-10.000
-12.000
-14.000
1.350E-06 1.400E-06 1.450E-06 1.500E-06 1.550E-06 1.600E-06
Wavelength [m]
(a) (b)
Fig. 12. Electric field distribution at (a) k ¼ 1:451 ½lm and (b) k ¼ 1:561 ½lm.

S3 ¼ 90:068½%, the distribution ratio of PORT 2 is S2 ¼ 5:297½%, and the distribution
ratio of PORT 1 is S1 ¼ 4:634½%. The electric field distribution is shown in Fig. 12.
The extinction ratio is 12:305 ½dB.
According to the results of this research, it was confirmed that dispersive linear and
nonlinear optical duplexer can switch the route according to the wavelength of the
input optical signal. By changing the structure of dispersive linear and nonlinear optical
duplexer, the distribution ratio, the extinction ratio, and the wavelength difference of
switching the route changed.
It is necessary to narrow the wavelength difference of switching the route, and
increase the numerical value of the extinction ratio.
References
1. Joannopoulos, J.D., Meade, R.D., Winn, J.N.: Photonic Crystals. Princeton University Press,
New Jersey (1995)
2. Noda, S., Baba, T. (eds.): Roadmap on Photonic Crystals. Kluwer Acedemic Publishers, The
Netherlands (2003)
3. Sullivan, D.M.: Nonlinear FDTD formulations using Z transforms. IEEE Trans. Microw.
Theory Tech. 43(3), 676–682 (1995)
4. Yee, K.S.: Numerical solution of initial boundary value problems involving Maxwell’s
equation. IEEE Trans. Antennas Propag. 14(3), 302–307 (1996). IEEE Trans. EMC 32(3),
222–227 (1990)
5. Li, X., Maeda, H.: Numerical analysis of photonic crystal directional coupler with kerr-type
nonlinearity. In: Proceedings of 5th Asia-Pacific Engineering Research Forum on Microwaves
and Electromagnetic Theory, pp. 165–168 (2004)
Cycle Interference Alignment for the Full
Duplex Communication System Based
on User Virtual Splitting Thought
Haiying Ren(&), Di Wu, Man Li, and Tianyi Feng
China Academy of Information and Communication Technology (CAICT),

Beijing 100191, China
renhaiying@caict.ac.cn
Abstract. This paper mainly focuses on the single-cell multi-user full-duplex

system. Both the base station and the users adopt the full duplex (FD) technol-
ogy. The interference alignment (IA) is introduced in the system to resolve the
interference from the intra-user and the inter-users. Based on the thought of the
user virtual splitting, the FD-BS-FD-user system is equivalently processed.
Then, a set of cycle IA equations are designed to reach the optimal system
degree of freedom (DoF). The simulation shows that the cycle IA scheme can
reach the desired result.
1 Introduction
The mobile communications for the future has become a global hot spot of research and
development. The efficient use of spectrum resources is the main goal in the design of
communication system. The new duplex technology provides the possibility to achieve
this goal. Particularly, the full duplex (FD) technology, which refers to the simulta-
neous operation of the transmitter and receiver at the same frequency, is an accepted
way to improve the spectrum efficiency of the system. In theory, the spectral efficiency
of the FD wireless communication systems can be doubled [1]. However, when the
technology of the FD is used in the communication system, there is a lot of the
additional interference. For example, the uplink users will produce interference in the
downlink users. If this interference is not canceled, the FD system cannot obtain two
times the spectral efficiency of the half duplex (HD) system [2].
At present, the research results for single-cell single-user full-duplex system are
relatively perfect [3]. However, for the single-cell multi-user full-duplex system, the
relative research is not quite comprehensive [4]. In the paper [4], it supposes that the
transmitting user does not interfere with the receiving user. Obviously, the result is too
optimistic. In addition, the complexity of the scheme is very high, especially when the
number of antennas in the node is relatively large. Therefore, the present algorithm is
not suitable for the actual system.
In recent years, the scholars have found other interference management mecha-
nisms besides interference elimination and interference orthogonalization. The most
representative of them is the interference alignment (IA) [5, 6]. The basic idea of IA is
to design the space of sending signals, and all the interference signals of each receiver

https://doi.org/10.1007/978-3-030-15035-8_27
Cycle Interference Alignment for the Full Duplex Communication System 287
are compressed into the smallest subspace. In this way, the larger signal space is
reserved to receive the target signal [7]. The degree of freedom (DoF) is the first
approximation of system data rate [8]. It is an important index to measure IA tech-
nology. Its physical meaning is the number of data flows that the system can transmit at
the same time. From the perspective of DoF, we can see that IA can make the capacity
of the system increase linearly with the number of users. It proves that the capacity of
multi-user system is not limited by interference. This has broken the past cognition that
the wireless communication system is the interference limited system. Therefore, IA
has aroused extensive attention and research in the academic circle.
In this paper, the IA technology is explored to solve the interference from the intra-
user and the inter-users in the single-cell FD-BS-FD-user system. Firstly, the user
virtual splitting thought is introduced, and the system is equivalent to the FD-BS-HD-
user system. Then, a set of cycle IA equations are designed to reach the optimal system
degree of freedom (DoF). The simulation shows that the cycle IA scheme can reach the
desired result.
2 The System Model
The system model of the single-cell multi-user full-duplex system is shown as Fig. 1.
The system parameters configured is (M, N, K). Both the base station (BS) and users
work in duplex mode. There are M antennas in the BS and N antennas in each user.
There are K users. Obviously, for each users, there are interference from inter-users, as
well as interference from the uplink signal of the intra-user. Gi,j is the interference
channel from user j to user i.
Since the base station cannot support M or more users in the same time, the
parameters in this system are supposed as M > K. In addition, we suppose the
parameters of (M, N, K) can satisfy the feasibility condition of IA [9].
3 The Cycle Interference Alignment Algorithm Based

on User Virtual Splitting Thought
For the FD-BS-FD-user system, to deal with interference problems, we adopt the cyclic
interference alignment algorithm based on the user virtual splitting thought. Firstly,
each full-duplex user is equivalently split into one virtual uplink user and one virtual
downlink user. Then, the cyclic interference alignment algorithm is adopted to elimi-
nate the interference from the intra-user and the inter-users. Finally, the whole system
reach the maximum DoF, that is, the maximum system capacity.
3.1 User Virtual Splitting Thought

As shown in Fig. 2, each actual full-duplex user is equivalent to two virtual half-duplex
terminals. One is the virtual uplink user and the other is the virtual downlink user. In
other word, the full-duplex user k is equivalent to half-duplex uplink user k and the
half-duplex downlink user k. Now, the FD-BS-FD-user system is equivalent to the FD-
288 H. Ren et al.
FD BS with M tx/rx antennas

tx as
ant nn
enn t e
as rx an
H i[ u ] H i[ d ] H j[ d ] H j[u ]
(i, j = 1, , K)
G i, j
G j ,i
K FD users with N tx/rx antennas

Downlink uplink inter-user
signal signal interference signal
Fig. 1. The system model
BS-HD-user system where K[u] = K[d] = K. In this way, for each virtual downlink user,
the inference signals come from all the K virtual uplink users.
3.2 Cyclic Interference Alignment Algorithm

For the equivalent FD-BS-HD-user system, the cycle IA is used in this system to cancel
the interference from the virtual uplink users. In this way, the optimal DoF can be
reached.
½u
The precoding matrix of the i virtual uplink user is V i , and the precoding matrix
½u
of the i downlink user is V i . The received beamforming vector of the i downlink user
½u
is W i .
For each downlink user, there are N spatial dimensions. In order to obtain at least
one DoF, all the interference from the uplink users should be aligned into the N-1
spatial dimensions.
In order to obtain the most optimal DoF, all the uplink users are coded according to
the follow loop alignment equations.
uplink signal
interference tx antennas rx antennas
1 N 1 N
tx antennas rx antennas .. .. tx antennas rx antennas
1 N 1 N 1 N 1 .. N
.. .. ..
...
UE 2
UE 1 UE K
inter-user interference signal
1 N 1 N
.. ..
virtual uplink virtual downlink

user 1 user 1
..
..
user 2 user 1
...
...
.. ..
user 3 user 1
Fig. 2. The user virtual splitting model
½u ½u
G1;1 V 1 ¼ t1 G1;2 V 2
½u ½u
G2;2 V 2 ¼ t2 G2;3 V 3
...
½u ½u
GK1;K1 V K ¼ tK1 GK1 ;K V K
½u ½u
GK;K V K ¼ tK GK;1 V 1
½u ½u
For each downlink user j, the interference space is spanðGj;1 V 1 ; Gj;2 V 2 ; . . .;
½u
Gj;K V K Þ.According to the feasibility condition of IA, when K[u] = K[d], the minimal
configuration of N should satisfy Nmin = max{K[u], 2}. Therefore, we can get
½u ½u ½u
rankðGj;1 V 1 ; Gj;2 V 2 ; . . .; Gj;k V K Þ ¼ K ½d 1\N. In this way, every downlink user
can receive one DoF at least without interference.
290 H. Ren et al.
In Fig. 3, we show the simulation results of the sum rate of the FD-BS-FD-user system
in different system configuration. It can be seen that the sum rate increases with the
number of the users. In other words, the cycle IA scheme can reach the desired results.
70 K=5
K=4
60 K=3
Average Sum Rate [bps/Hz]
50
40
30
20
10
0
0 5 10 15 20 25
SNR [dB]
Fig. 3. Sum rate versus different system configurations
In addition, for the single-cell multi-user FD-BS-FD-user system, although the

interference increases with the users, the system does not become the interference
limited system. On the contrary, by using the cycle IA, the sum rate increases with the
number of the users.
In Fig. 4, we show the simulation results of the sum rate of the single-cell multi-
user FD-BS-FD-user system of the cycle IA and the conventional scheme. For the
conventional scheme, the BS uses the zero-forcing (ZF) precoder/combiner, and each
user is assumed to use maximum ratio combining (MRC). It is observed that the
proposed cycle IA exhibits higher sum-DoF compared to the conventional scheme.
70 Cycle IA
ZF-MRC
60
Average Sum Rate [bps/Hz]
50
40
30
20
10
0
0 5 10 15 20 25
SNR [dB]
Fig. 4. Sum rate versus different algorithms
5 Conclusions
In this paper, we focus on the single-cell multi-user full-duplex system. By using the
cycle IA, the sum rate of the system increases with the system parameter configuration.
The method in this paper can be promoted to the case that the number of the user data
rate more than one. Furthermore, it can reach the maximum sum-DoF, and satisfy the
transmission requirements of system uplink and downlink.
Acknowledgments. This work was supported by the Beijing Natural Science Foundation
(4172071), Youth Project of China Academy of Information and Communications Technology.
References
1. Sabharwal, A., Schniter, P., Guo, D., Bliss, D.W., Rangarajan, S., Wichman, R.: In-band full-
duplex wireless: challenges and opportunities. IEEE J. Sel. Areas Commun. 32(9), 1637–
1652 (2014)
2. Goyal, S., Liu, P., Panwar, S., DiFazio, R.A., Yang, R., Bala, E.: Full duplex cellular systems:
Will doubling interference prevent doubling capacity? IEEE Commun. Mag. 53(5), 121–127
(2015)
292 H. Ren et al.
3. Nguyen, D., Tran, L.-N., Pirinen, P., Latva-aho, M.: Precoding for full-duplex multiuser
MIMO systems: Spectral and energy efficiency Maximization. IEEE Trans. Signal Process. 61
(16), 4038–4050 (2013)
4. Nguyen, D., Tran, L.-N., Pirinen, P., Latva-aho, M.: On the spectral efficiency of full-duplex
small cell wireless systems. IEEE Trans. Commun. 13(9), 4896–4910 (2014)
5. Cadambe, V.R., Jafar, S.A.: Interference alignment and degrees of freedom of the, K-user
interference channel. IEEE Trans. Inf. Theory 54(8), 3425–3441 (2007)
6. Kim, D., Torlak, M.: Timization of interference alignment beamforming vectors. IEEE J. Sel.
Areas Commun. 28(9), 1425–1434 (2010)
7. Sahai, A., Diggavi, S., Sabharwal, A.: On degrees-of-freedom of full-duplex uplink/downlink
channel. In: Proceedings of the IEEE Information Theory Workshop, Seville, Spain (2013)
8. Jafar, S.A., Fakhereddin, M.J.: Degrees of freedom for the MIMO interference channel. IEEE
Trans. Inf. Theory 53(7), 2637–2642 (2007)
9. Kim, K., Jeon, S.W., Kim, D.K.: The feasibility of interference alignment for full-duplex
MIMO cellular networks. IEEE Commun. Lett. 19(9), 1500–1503 (2015)
The 12th International Symposium on
Mining and Web (MAW-2019)
Proposal of Web API Architecture
for Smart Community: A Case Study
of Japan
Kanae Matsui(B)
Tokyo Denki University, 5 Senju Asahi-cho, Adachi-ku, Tokyo 120-8551, Japan

matsui@mail.dendai.ac.jp
Abstract. This paper proposes a Web API (Application Programming

Interface) design for IoT data collected from a smart community’s plat-
form. The IoT data of a smart community should be utilized for residents,
who are owners of the data, for services and applications because storing
data makes cost. Therefore, a smart community would have a database
that collects the IoT data, and a Web API that distributes only allowed
the data for third party, which would provide smart community services
to the community, is necessary. To provide IoT services and applica-
tions to a smart community, IoT data, which is specifically collected by
IoT devices, should be distributed after analyzing data meaning. Thus,
the Web API provides both IoT data and meaning to enhance devel-
oping applications such as high-dense temperature information. As a
case study, we show the smart-community-based Web API first, and
then show examples on the utilization of real IoT data collected in
Urawamisono, Saitama city, which is approximately 30 km from central
Tokyo, in Japan.
1 Introduction
A smart community consists of numerous Internet of Things (IoT) devices and
handles massive IoT data for providing data-driven services to residents who are
owners of the data. Setting up IoT devices in a community involves high cost, and
making benefits from IoT data, which are collected by IoT devices, is necessary
to manage a smart-community-based IoT platform. Therefore, the IoT platform
should have a Web API (Application Programming Interface) to provide IoT
data to a third party that provides IoT-based services and applications. However,
only providing IoT data is not enough to enhance their business, because the
extraction of meaningful information from IoT data requires an understanding
of the background.
From this context, this study proposes a Web API system and utilizes a col-
lected by IoT devices that are set in Urawamisono, Saitama city, Japan, where is
proceeding to be smart community [1]. In addition, the system puts background
information, which needs to understand data meaning, into the data. In this
time, indoor temperature is utilized as a sample to show a way how to utilize
https://doi.org/10.1007/978-3-030-15035-8_28
296 K. Matsui
and evaluation, because indoor temperature is related to many services, e.g.,

maintaining indoor comfort and health to prevent indoor heat shocks [2].
2 Related Work
Two software, including a Web API that handles IoT for data-driven systems,
which consist of a smart city as their component, are introduced. Our target is
a “smart community,” which is smaller than a city, but the target is the same in
terms of IoT data utilization. In this section, representative Web API functions
and differences from the proposed Web API.
2.1 FIWARE: Open Platform for Smart City

FIWARE is software that aims to distribute IoT data for IoT-based services
such as sightseeing navigation, face authentication settlement, traffic congestion
avoidance, crime prevention measures, and housing services, developed by NEC
Corporation [3]. This service is the environment for the verification of data-
utilization infrastructure service that a smart city provides to its residents.
FIWARE is a data management infrastructure that focuses on cross-sectoral
data distribution. They have seven categories, approximately 40 types of mod-
ules, can be freely combined, and can be used for the application. Each module
is defined by “NGSI (Next Generation Service Interface)” standardized by OMA
(Open Mobile Alliance), and data is exchanged through this [4].
2.2 CityTM IoT Platform

A cityTM IoT Platform captures and processes high-density sensor data from the
ubiquitous CityIQ’s intelligent nodes, developed by General Electric Company,
installed on all street lights across cities [5]. In addition, it collects both real-time
and historical use cases. This platform provides enables app developers to access
value-add data derived from rich analytics running on CityIQ’s intelligent nodes
and on the cloud for a third party that wants to utilize IoT data.
This platform has five main APIs, viz., parking planning, pedestrian plan-
ning, situational awareness, environmental planning, and traffic planning. With
these APIs, developing IoT-based services would be much easier.
2.3 Smart Community Project in Urawamisono

We have conducted a smart community project, named Smart Community
Infrastructure Platform (SCIP), which involved the installation of IoT devices
in households and has collected the data in Urawamisono. Urawa Misono is a
subcenter of Saitama City, the most populous city in the Saitama Prefecture,
Japan. While Saitama City boasts a population of 1.26 million, Urawamisono
has remained largely rural. Only 45 min from Tokyo by rail, it is the final stop on
the Saitama Rapid Railway Line. Saitama Stadium is an important site for the
Proposal of Web API Architecture for Smart Community 297
2020 Summer Olympic Games, prompting local and regional officials to consider
how they will accommodate the massive influx of event spectators and maximize
the impact of this influx for broader development goals.
This paper proposes a Web API for this developing smart community, which
has the following differences from above-mentioned smart-city APIs. The above-
mentioned APIs were designed for targeted services, e.g., sightseeing navigation,
traffic congestion avoidance, and housing services. We also aim to distribute IoT
data for third parties that would like to conduct such services by using IoT data.
Interrupting IoT data requires expert knowledge for each data, but it takes a
cost. Therefore, it is necessary to add interrupting data into distributed IoT
data, e.g., temperature prediction. Therefore, the proposed Web API distributes
interruption information into IoT data to enhance third party developing smart
community services.
3 System Description
In this section, a web API system for IoT data utilization based on a smart
community is described. Based on the SCIP project, the SCIP database, which
stores all IoT data in the community, has an interface to access the data from
the third party. The SCIP database has collected various types of IoT data, e.g.,
indoor temperature, humidity, CO2 emission, luminance, PM2.5, etc. The indoor
temperature and humidity can be used for many services, such as smart house
functions, which maintain indoor comfort with an automatic HVAC (Heating,
Ventilation, and Air Conditioning) system. Therefore, the proposed API targets
both indoor temperature and humidity data in this case. Figure 1 illustrates a
conceptual diagram of the Web API.
Fig. 1. Conceptual diagram of Web API.
This Web API has a function to add expert knowledge such as appropriate
value and influence on human body as interpretation information when utiliz-
ing IoT data gathered in the smart community and making it into service. In
this research, in order to deal with This Web API has a function to add expert
298 K. Matsui
Table 1. System development environment.
Purpose Programming language (version)

Applications Ruby (ver.2.2.2)
Framework and database Sinatra (ver.2.0.0)
Database PostgreSQL d(ver.10.0)
knowledge, such as appropriate value and influence on human body, as interpre-

tation information when utilizing IoT data gathered in the smart community
and turning it into a service. In this research, in order to deal with temperature
data, we investigate the appropriate temperature for each season and create a
conditional branch. An interpretation sentence is determined by this conditional
branch, and this is combined with the IoT data. The appropriate temperature
for each season is given in Table 4 below. For spring and autumn, which are
not included in the table, the appropriate temperature was set to 21–23 ◦ C by
using indices of indoor proper temperature. Interpretation data is created by a
program different from the Web API. IoT data and date data are converted into
interpreted sentences by conditional branching in the interpretation program.
As a premise of this system, the proposed system is in the smart community
database, where the data of IoT devices are stored. The system adds interpre-
tation information to IoT data and converts it into JSON (JavaScript Object
Notation) format data for digital distribution. The system consists mainly of
three functional requirements, viz., (1) access to the smart community database,
(2) biding of IoT data and interpretation data, and (3) conversion to distribution
data. Figure 2 illustrates an overview of the system and Table 1 gives the system
development environment.
Fig. 2. Overview of data collection.

3.1 Access to the Smart Community Database

For data access of this Web API, (1) to designate the building ID and sensor type
or to acquire all data, (2) to specify the building ID and sensor type, the data
acquisition period to acquire data, (3) to designate the ID of each building and
the sensor type, and acquiring the data closest to the current time. The building
denotes a smart house in a smart community. The data can be acquired by
adding parameters according to the purpose to the base URI (Uniform Resource
Identifier). Table 2 lists the parameters for each function as an example.
Table 2. Example of request parameter.
Request Parameter
All temperature data of /simple?sensor=temp+home id=1
Building ID 1
From 2017−2−25 to current /custom?sensor=temp+home id=1
Building ID 1
All temperature data up to start=2017−2−25+end=2017−10−3
2017−10−3
Temperature data acquired /now?sensor=temp+home id=1
at the time closest to the
current time of building ID 1
3.2 Biding of IoT Data and Interpretation Data

Each IoT data collected from a household in the smart community was accumu-
lated and stored in the smart community database. Although various types of
IoT data are stored in the smart community database, indoor temperature data
is selected in this demonstration. Table 3 depicts a data frame of the system.
Table 3. Data frame of the proposed system.
Name of column Type Contents

ID Integer Serial number
Date Date Date of IoT data acquisition
(YYYY-MM-DD)
Temp Double precision IoT data value
The proposed Web API has the function of attaching expert knowledge, such
as appropriate value and influence on human body, as interpretation information
when utilizing IoT data gathered in the smart community and turning it into
300 K. Matsui
a service by the third party as mentioned previously. In order to deal with the
indoor temperature data, the system investigates the appropriate temperature
for each season and creates a conditional branch. An interpretation sentence is
determined by this conditional branch, and this is combined with the IoT data.
The appropriate temperature for each season is given in Table 4. The conditional
branch is defined by the Japanese season.
Table 4. Response table in proposed Web API.
Season Seasonal classification Proper indoor temperature

Summer From June to August 26–27◦
Winter From December to February 20–22◦
Intermediate period Others 22–23◦
The proper indoor temperature in Table 4 is defined as suitable interpretation

data. IoT data are converted into interpreted sentences by conditional branching
in the interpretation program installed in the web API database.
3.3 Conversion to Distribution Data
In order to distribute IoT data and interpretation data as digital data, the
distribution data is defined in the JSON format. The response fields are given
in Table 5.
Table 5. Response field.
Property name Contents

ID Serial number
date date IoT data acquisition date (YYYY-MM-DD)
time IoT data acquisition time (HH:MM:SS)
temp IoT data value
temptxt Interpretation
4 Evaluation
In order to verify the system, dummy data of indoor temperature was registered
in the constructed database. For confirmation, to start the browser on the client
server connected to the same network as the application server, and input the
URI according to the purpose in the address bar. If the response is correct,
displaying the data in the JSON format would be on the browser of the client
server.
In a Web API, a response time for the third party is necessary to provide IoT
data. In order to verify the response time of this Web API, Postman (ver.5.5.0)
is used [6]. Postman is a Web API testing tool and used for measurement of the
response time. Table 6 lists the information of equipment used for the evaluation.
Table 6. Equipment list.
Model Role OS Serial number IP address

number
Mac book Application mac OS C02N71THG3QR 192.168.2.45
Pro server for (High Sierra)
web API
HPCT Client Windows 10 00179-68585- 192.168.2.9
M116dm3-G server Pro 52662-AAOEM
As a test, three acquisition methods are executed 30 times by the client server
and the result is verified.
1. Acquiring all data by specifying the building ID and sensor type

2. Specifying the ID of the building and sensor type
3. Specifying the ID of the building and the sensor type and acquiring the data
closest to the current time
When acquiring data by each method, it shows the URI entered in the address
bar of the browser on the client server and displays the result in a web browser.
5 Results
5.1 Item 1
The item 1 is to check when user sent a request to proposed web API with spec-
ifying the building ID and sensor type to acquire all data. Following URI is a
request to acquire all data. The URI includes that IP address, sensor type, and
building ID.
http://192.168.2.45:4567/simple?sensor=temp&home/ id=1
Accordingly, all data-acquisition methods of this Web API were verified and
it was confirmed that the results were displayed correctly. The result of JSON
files in a web browser is shown in Fig. 3.
302 K. Matsui
Fig. 3. Displaying results of item 1.
5.2 Item 2
The item 2 is to check when user sent a request to proposed Web API with
specifying the ID of the building and sensor type. The URI for data request
consists of IP address, sensor type, building ID, which means home ID in this
case, and data when users want to acquire. Figure 4 shows the result of this
request, and then the result showed all requested data were acquired from the
proposed Web API by this type of requests.
5.3 Item 3
The item 3 is to check when user sent a request to proposed Web API with
specifying the ID of the building and the sensor type and acquiring the data
closest to the current time. Following URI is a request to acquire the most
current data.
http://192.168.2.45:4567/now?sensor=temp& home id=1
Figure 5 shows the result of this request, and then the result showed all
requested data were acquired from the proposed Web API by this type of
requests.
5.4 Response Time

The response time The average value of 30 response-time measurements by using
Postman is the response-time verification result. The average of the response
time was 26.5 s, including acquiring all data.
In conclusion, the proposed Web API system results showed that all requests
to acquire IoT and interpretation data were completed. However, the response
time took a time more than a request from some services and applications. As
the next step, the response time of the system should be improved to accept all
requests from databases of third parties.
6 Conclusion
In this research, we proposed and implemented the Web API as a data access
interface for creating services and applications using IoT data accumulated in a
smart community. We constructed a database and Web API that are supposed to
be used in the smart community and evaluated this Web API in terms of validity
and response time. Accordingly, in validation verification, we were able to receive
the correct response to the request according to the purpose. In response-time
measurement, it was a Web API with high immediate responsiveness.
From this result, it was thought that it takes more time to acquire data when
operating in the actual smart community. Therefore, in order to use enormous
IoT data by using the Web API, there is a need to design with consideration of
functionality and response time. In this research, the design was realized with
the method of acquiring all specified sensor data, but by making it possible to
set the upper limit of data acquisition number, it is considered that the response
time could be shortened without impairing the functionality.
Acknowledgements. This work was supported in part by the R&D project “Design
of Information and Communication Platform for Future Smart Community Services”
by the Ministry of Internal Affairs and Communications (MIC) of Japan, and was sup-
ported by Funds for the Integrated Promotion of Social System Reform and Research
and Development, MEXT, Japan, and by MEXT/JSPS KAKENHI Grant (B) Number
17K12789 and SECOM Science and Technology Foundation.
References
1. Miura, T., Wijekoon, J.L., Prageeth, S., Nishi, H.: Novel infrastructure with common
API using docker for scaling the degree of platforms for smart community services.
In: 2017 IEEE 15th International Conference on Industrial Informatics (INDIN),
pp. 474–479. IEEE (2017)
2. Kanae, M.: Pleasant sleep provision system during warm nights as a function of
smart home. In: IEEE 2017 International Symposium on Networks, Computers and
Communications (ISNCC), pp. 1–6. IEEE (2017)
304 K. Matsui
3. Erno, K., Martin, B., Jaeho, K., et al.: Standards-based worldwide semantic inter-
operability for IoT. IEEE Commun. Mag. 54(12), 40–46 (2016)
4. Ernoe, K., Martin, B.: Apparatus and method for connecting at least two systems
by converting data. Google Patents, US Patent App. 15/742,069 (2018)
5. General Electric Company, CityTM IoT Platform. https://developer.currentbyge.
com/cityiq. Accessed 25 Jan 2019
6. Postman, Inc. Postman. https://www.getpostman.com/. Accessed 25 Jan 2019
Evaluation Measures for Extended
Association Rules Based on Distributed
Representations
Tomonobu Ozaki(B)
College of Humanities and Sciences, Nihon University,

3-25-40 Sakurajosui, Tokyo 156-8550, Japan
tozaki@chs.nihon-u.ac.jp
Abstract. Indirect association rules and association action rules are

two notable extensions of traditional association rules. Since these two
extended rules consist of a pair of association rules, they share the same
essential drawback of association rules: a huge number of rules will be
derived if the target database to be mined is dense or the minimum
threshold is set low. One practical approach for alleviating this essential
drawback is to rank the rules to identify which one to be examined first
in a post-processing. In this paper, as a new application of representation
learning, we propose evaluation measures for indirect association rules
and association action rules, respectively. The proposed measures are
assessed preliminary using a dataset on Japanese video-sharing site and
that on nursery.
1 Introduction
Frequent itemset mining [1,2], i.e., the enumeration problem of all itemsets
appearing frequently in databases is one of the most fundamental problems in
data mining. As similar, discovery of association rules that consist of two fre-
quent itemsets is the most traditional method in this field. Various extensions
of frequent itemsets and association rules in a wide variety of application fields
have been developed over a long period. Since an association rule consists of
two frequent itemsets, association rule discovery has the same drawback as the
frequent itemset mining has. That is combinatorial explosion. A huge number
of itemsets and rules will be derived if the target database to be mined is dense
or the minimum threshold is set low. To suppress insignificant results, several
techniques have been proposed such as mining condensed representations [3,4]
and filtering based on statistical tests [5,6]. In addition, to rank the results based
on certain viewpoints is also useful in practice, a number of evaluation measures
have been proposed [7,8] including Lift [9] and Conviction [10]. Especially in
recent years, various evaluation measures based on distributed representation
[11] have been proposed. For example, some methods are proposed [12,13] to
evaluate the degree of representatively, exceptionality and significance of fre-
quent itemsets based on the distances among frequent itemsets in a vector space
https://doi.org/10.1007/978-3-030-15035-8_29
306 T. Ozaki
obtained using existing techniques on representation learning. Another method

[14] evaluates frequent itemsets and association rules locally using distributed
representations of their components, i.e., items in frequent itemsets and itemsets
in association rules.
As an extension of the association rule, two special rules were proposed.
One is indirect association rules (IARs in short) [15] which capture a contrast
relation under a certain condition. The other is association action rules (AARs
in short) [16] which express the relation on the change between attribute values.
Although IARs and AARs are sophisticated rules expressing useful knowledge
in practice, because they formally consist of two association rules, they inherit
the same essential drawback of association rules. To overcome the drawback, an
evaluation measure for IARs in the domain of cooking recipes [17] and one for
AARs which considers constraints and costs [18] were proposed, respectively.
However, because these two measures require external domain knowledge, they
are not easy to use in a wide range of applications.
From the background explained above, in this paper, we propose evaluation
measures not requiring external knowledge for IARs and AARs based on dis-
tributed representations.
The rest of this paper is organized as follows. After the formal definitions of
IARs and AARs are given as preliminary in Sect. 2, the evaluation measures for
them based on distributed representations are proposed in Sect. 3. The results of
preliminary experiments are reported in Sect. 4. Finally, we conclude the paper
and describe future directions in Sect. 5.
2 Preliminary
Let I = {i1 , i2 , · · · , i|I| } be a set of all items. A transaction t ⊆ I is a set of
items in I, and a database D = {t1 , t2 , · · · , t|D| } is a set of transactions. A
pattern P ⊆ I is a combination of items which appears in plural transactions in
common. The support value of P with respect to D is formally defined as
supD (P ) = | {t ∈ D | t ⊆ P } | / |D| .
Given the minimum support threshold σ (0 < σ ≤ 1), if a pattern P satisfies

the condition supD (P ) ≥ σ, then P is said to be a frequent pattern. We use the
terms frequent patterns and frequent itemsets interchangeably. The purpose of
frequent itemset mining is to derive the whole set
σ
FD = {P ⊆ I | supD (P ) ≥ σ}.
of frequent patterns in D.
For two frequent patterns A and C having no item in common (A ∩ C = ∅),
we consider a rule in the form of A → C representing that if A holds then C also
holds. The support and confidence values of the rule A → C are respectively
defined as
supD (A → C) = supD (A ∪ C) and confD (A → C) = supD (A → C)/supD (A).

Evaluation Measures for Extended Association Rules 307
Given the minimum support threshold σ (0 < σ ≤ 1) and the minimum

confidence threshold θ (0 < θ ≤ 1), if a rule A → C satisfies the both conditions
supD (A → C) ≥ σ and confD (A → C) ≥ θ, then the rule is called an association
rule. The problem of association rule discovery is to derive the whole set
⎧ ⎫
⎨ A, C ⊆ I, A ∩ C = ∅, ⎬

RD = A → C supD (A → C) ≥ σ,
σ,θ
⎩ confD (A → C) ≥ θ ⎭
of association rules in D with respect to σ and θ.
The indirect association rules [15] (IAR in short) are the pairs of association
rules representing a contrast relation under a certain condition. Given the min-
imum support threshold σmin (0 < σmin ≤ 1), the maximum support threshold
σmax (0 ≤ σmax ≤ 1) and the minimum dependency threshold τ , if a frequent
pattern M ⊆ I and two items x and y ∈ I satisfy the condition
supD (M → {x}) ≥ σmin ∧ dD (M, {x}) ≥ τ ∧
supD (M → {y}) ≥ σmin ∧ dD (M, {y}) ≥ τ ∧
supD ({x, y}) < σmax
where dD (X, Y ) is a function such as confidence and lift to measure the depen-
dency between X and Y in D, then the pair of two association rules
M → {x}, M → {y}
is called an indirect association rule.
The association action rules [16] (AAR in short) are the pairs of association
rules representing that the changes in antecedents under certain conditions bring
the changes in consequences. In other words, they are the rules that expresses
the explicit behaviors, or changes in antecedents, required to change the outcome
to the desired state.
The association action rules are defined on attribute-value tables. The whole
set of attributes A = As ∪ Af consists of a set As of stable attributes which
cannot change the value such as age and a set Af of flexible attributes which
permit the change of value such as jobs. We denote the range of an attribute
a ∈ A as Va and assume all values v ∈ Va are discrete. By constructing an item
in the form of a : v for an attribute a ∈ A and its value v ∈ Va , we can convert
each row in an attribute-value table into a transaction
{a1 : va1 , a2 : va2 , a3 : va3 , · · · , a|A| : va|A| }.
An attribute-value table can be regarded as a transaction database by this
conversion. For a pattern P = {aP1 : vP1 , · · · , aP|P | : vP|P | }, three notations
AP = {aP1 , · · · , aP|P | }, AP P P P
s = A ∩ As and Af = A ∩ Af are prepared
to denote the set of all, stable and flexible attributes in P , respectively. If two
association rules r1 = P1 → Q1 and r2 = P2 → Q2 satisfy the following five
conditions, then the rule
r1 ⇒ r2 (≡ (P1 → Q1 ) ⇒ (P2 → Q2 ) )
is said to be an action rule.
308 T. Ozaki
1. AP1 = AP2 : Two antecedents consist of the same set of attributes.

2. ∀a ∈ APs [ a : v1 ∈ P1 ∧ a : v2 ∈ P2 ∧ v1 = v2 ]: All the values of the stable
1
attributes in the antecedents match between two rules.

3. ∃a ∈ APf [ a : v1 ∈ P1 ∧ a : v2 ∈ P2 ∧ v1 = v2 ]: Two antecedents have at least
1
one flexible attribute with different values.

4. AQ1 = AQ f
1
= AQ2 = AQ 2
f : Two consequences consists of the same flexible
attributes.
5. ∀a ∈ AQ1 [ a : v1 ∈ Q1 ∧ a : v2 ∈ Q2 ∧ v1 = v2 ]: No attribute having the
same value exists in two consequences.
Given an action rule r1 ⇒ r2 , the support and confidence values are respec-
tively defined as follows [16].
supD (r1 ⇒ r2 ) = min (supD (r1 ), supD (r2 ))

confD (r1 ⇒ r2 ) = confD (r1 ) · confD (r2 )
Given the minimum support threshold σ (0 < σ ≤ 1) and the minimum con-
fidence threshold θ (0 < θ ≤ 1), if an action rule r1 ⇒ r2 satisfies the conditions
supD (r1 ⇒ r2 ) ≥ σ and confD (r1 ⇒ r2 ) ≥ θ, then it is called an association
action rule.
3 Evaluation Measures for Extended Association Rules
3.1 Distributed Representations of Items and Patterns
Representation learning [11] can associates low dimensional real vectors to

objects. The obtained vector representation is called distributed representation.
In this paper, we build evaluation measures for extended association rules using
→
− →
−
the distributed representations i for an item i ∈ I and P for a pattern P ⊆ I.
It is assumed that the distributed representations of items can be obtained by
existing methods such as GloVe [19], word2vec [20,21] and fastText [22]. Using
→
−
smooth inverse frequency weighting [23], the distributed representation P of a
pattern P ⊆ I is derived as
−
→ →
−
P = i∈P wi i / i∈P wi
where wi = a/(a + supD ({i}). In this paper, the parameter a is set to 1. As a

distance between two objects Oi and Oj , we employ a normalized cosine distance
−
→ −→
between corresponding vectors Oi and Oj formally defined as

−
→ − → 1
−
→ − →
dist Oi , Oj = 1 − cos(Oi , Oj ) .
2
3.2 Evaluation Measures for Indirect Association Rules
An indirect association rule M → {x}, M → {y} intuitively claims that x and

y play the similar role under the condition M . Therefore, in the evaluation of
IARs, it is essential to consider the degree of relevance between x and y with
respect to M . According to this basis, we propose three evaluation measures
shown below.

−−−−−→ −−−−−→
EIgain (M, x, y) = dist ( −
→
x, →
−
y ) − dist M ∪ {x}, M ∪ {y}

−−−−−→
−−−−−→
EIavg (M, x, y) = 1
2 dist →−
x , M ∪ {y} + dist −→
y , M ∪ {x}

→
−
→
−

EIavg (M, x, y) = 1
2 min dist →
−
x, i + min dist →
−
y, i
i∈M ∪{y} i∈M ∪{x}
The measure EIgain represents the difference in distance between the case of
considering the condition M and that of not considering M . It gives high value
for rules if the distance between x and y becomes smaller in the consideration of
M . The idea behind EIavg is that the indirect association rules are worthwhile
if it is difficult to guess the consequences of one rule from the other rule. By
regarding the distance as the degree of difficulty for the estimation, the measure
EIavg employs the average value of difficulties between the case for estimating x

from M ∪ {y} and that for estimating y from M ∪ {x}. The measure EIavg is a
variant of EIavg , and it measures the difficulty for estimating the consequences
by using each item in the rules separately.
3.3 Evaluation Measures for Association Action Rules
An association action rule r1 ⇒ r2 ≡ (P1 → Q1 ) ⇒ (P2 → Q2 ) intuitively rep-

resents the situation where the changes from P1 to P2 in the antecedents derive
the changes from Q1 to Q2 in the consequences. Therefore, we have to consider
changes in both of antecedents and consequences at a time in the evaluation of
the rules. In this paper, we believe that the rules are worthwhile if small changes
in antecedents lead to major changes in consequences. As an evaluation measure
reflecting this basis, we propose a measure

−→ − →
−→ −→
EAg (r1 , r2 ) = dist Q1 , Q2 − dist P1 , P2 .
In addition, using a set AP = AP P

s ∪ Af of common attributes in antecedents
and a set AQ of common attributes in consequence, we propose an evaluation
measure EAl as follows.
−−−−−−−−−→ −−−−−−−−−→
−−→ −−−→
EAl (r1 , r2 ) = dist −
a : v1 , a : v2 − dist AP P
s ∪ {a : v1 }, As ∪ {a : v2 }
a∈AQ ,a:v1 ∈Q1 ,a:v2 ∈Q2 a∈AP
f
,a:v1 ∈P1 ,a:v2 ∈P2
310 T. Ozaki
Both measures are based on the difference in distance between the change in
consequences and that in antecedents. They give a high value to an association
action rule whose change in consequences is large but that in antecedents is
small. While EAg evaluates the rules from more global viewpoint by captures
the changes at the level of itemsets, EAl determines the evaluation locally by
considering each attribute separately.
4 Experiments
4.1 Experiments for Indirect Association Rules
To assess the proposed evaluation measures for indirect association rules, we con-
ducted preliminary experiments using Niconico dataset1 provided by DWANGO
Co., Ltd. and National Institute of Informatics in Japan. It is a real dataset on
Japanese video-sharing site Nicovideo2 . In the experiments, each set of tags asso-
ciated with each movie posted in Nicovideo is employed as a transaction. From
the entire dataset, a database Dmusic having about 456 thousands of transac-
tions was prepared by collecting transactions having a tag “music”. Using the
confidence value as a dependency measure, we obtained 1,162 of indirect associ-
ation rules under the settings of σmin = 0.0001, σmax = 0.00001, and τ = 0.15.
As a distributed representation of items, we derive d-dimensional real vectors
where d ∈ {50, 150} using GloVe [19]3 .

To assess the proposed three measures EIgain , EIavg and EIavg , we cal-
culated the basic statistics of values of each evaluation measure for the set of
indirect association rules. The results are summarized in Table 1.
Table 1. The basic statistics of evaluation measures for indirect association rules
d = 50 d = 150

EIgain EIavg EIavg EIgain EIavg EIavg
Minimum value 0.050 0.089 0.095 0.067 0.086 0.104
First quartile 0.205 0.227 0.214 0.218 0.283 0.260
Median 0.261 0.277 0.254 0.286 0.323 0.305
Third quartile 0.327 0.316 0.298 0.343 0.359 0.342
Maximum value 0.489 0.486 0.470 0.489 0.481 0.473
Average 0.263 0.273 0.255 0.279 0.312 0.298
Standard deviation 0.087 0.071 0.063 0.083 0.070 0.068
From the results, it is observed that the measure EIgain has wider range and
larger standard deviation compared with other measures. Although the value of
1
https://www.nii.ac.jp/dsc/idr/nico/nico.html.
2
http://www.nicovideo.jp/.
3
https://nlp.stanford.edu/projects/glove/.
EIgain can become negative in theory, the minimum value exceeded zero in the
experiments. We believe that this results come from that the conditions M are
sufficiently considered at the stage of extracting indirect association rules.
Table 2. Correlation among evaluation measures
d = 50 d = 150
EIgain – EIavg 0.500 0.402

EIgain – EIavg −0.033 0.076

EIavg – EIavg 0.684 0.775
The Pearson correlations among proposed measures are shown in Table 2.

The positive correlations can be recognized between EIgain and EIavg as well as

EIavg and EIavg . It is confirmed that there exists certain relationships between
the similarity of consequences x and y under the condition M and the difficulty
of estimations for consequences.
4.2 Experiments for Association Action Rules
The Nursery Data Set4 was used in the experiments. This data contains 12,960
transactions with nine attributes. While three attributes are treated as stable
attributes, the remained six attributes are used as flexible. After deriving asso-
ciation action rules using σ = 0.0005 and θ = 0.70, we selected 41,727 rules
which have up to three flexible attributes in the antecedent, at least one stable
attribute in the antecedent and one flexible attribute in the consequence. The
distributed representation of items were derived by the same method as in the
case of the experiments for indirect association rules.
Table 3. The basic statistics of evaluation measures for association action rules
d = 50 d = 150
EAg EAl EAg EAl
Minimum value −0.624 −0.912 −0.661 −0.924
First quartile −0.170 −0.154 −0.187 −0.150
Median 0.037 0.098 0.043 0.114
Third quartile 0.121 0.141 0.132 0.157
Maximum value 0.433 0.439 0.439 0.444
Average −0.026 −0.001 −0.026 0.011
Standard deviation 0.180 0.204 0.191 0.208
4
https://archive.ics.uci.edu/ml/datasets/nursery.
312 T. Ozaki
We calculated the basic statistics of values of EAg and EAl for the set of
association action rules. The results are summarized in Table 3.
In the experiments, from the influence of the constraint on the form of associ-
ation action rules, most of rules had negative values irrespective of the evaluation
measures. It is recognized that the measure EAl has wider range and larger stan-
dard deviation. Although not shown in detail, a number of outliers were detected
in EAl . The correlation between EAg and EAl is very high. It is 0.908 for d = 50
and 0.918 for d = 150. In addition, the top ranked association action rules with
respect to the each evaluation measure were the same. From the results, no large
difference between two measures was observed unfortunately.
5 Conclusion
In this paper, to alleviate the essential drawback in association rule discovery, we

propose evaluation measures for indirect association rules and association action
rules using distributed representations. In addition, using the real world datasets,
the basic tendency of the proposed evaluation measures was investigated.
As future work, we have to evaluate the proposed measures from both the-
oretical and practical viewpoints using plural large scale datasets. We also plan
to utilize various techniques for obtaining distributed representations as well as
to extend the proposals to structured data mining.
Acknowledgements. In this paper, the author used the “Nicovideo dataset” pro-
vided by National Institute of Informatics. This work was partially supported by JSPS
KAKENHI Grant Number 17K00315.
References
1. Agrawal, R., Imielinski, T., Swami, A.: Mining association rules between sets of
items in large databases. In: Proceedings of the 1993 ACM-SIGMOD International
Conference on Management of Data, pp. 207–216 (1993)
2. Han, J., Cheng, H., Xin, D., Yan, X.: Frequent pattern mining: current status and
future directions. Data Min. Knowl. Disc. 15(1), 55–86 (2007)
3. Boulicaut, J.-F., Bykowski, A., Rigotti, C.: Free-sets: a condensed representation of
boolean data for the approximation of frequency queries. Data Min. Knowl. Disc.
7(1), 5–22 (2003)
4. Pasquier, N., Bastide, Y., Taouil, R., Lakhal, L.: Discovering frequent closed item-
sets for association rules. In: Proceedings of the 7th International Conference on
Database Theory, pp. 398–416 (1999)
5. Webb, G.I.: Discovering significant patterns. Mach. Learn. 68(1), 1–33 (2007)
6. Hämäläinen, W., Webb, G.I.: Statistically sound pattern discovery. In: Proceedings
of the 20th ACM SIGKDD International Conference on Knowledge Discovery and
Data Mining, p. 1976 (2014)
7. Wu, T., Chen, Y., Han, J.: Re-examination of interestingness measures in pattern
mining: a unified framework. Data Min. Knowl. Disc. 21(3), 371–397 (2010)
8. Lenca, P., Vaillant, B., Meyer, P., Lallich, S.: Association rule interestingness mea-
sures: experimental and theoretical studies. In: Guillet, F.J., Hamilton, H.J. (eds.)
Quality Measures in Data Mining, pp. 51–76 (2007)
9. Brin, S., Motwani, R., Silverstein, C.: Beyond market baskets: generalizing asso-
ciation rules to correlations. In: Proceedings of the 1997 ACM SIGMOD/PODS
Joint Conference, pp. 265–276 (1997)
10. Brin, S., Motwani, R., Ullman, J.D., Tsur, S.: Dynamic itemset counting and impli-
cation rules for market basket data. In: Proceedings of the 1997 ACM SIGMOD
International Conference on Management of Data, pp. 255–264 (1997)
11. Bengio, Y., Courville, A., Vincent, P.: Representation learning: a review and new
perspectives. IEEE Trans. Pattern Anal. Mach. Intell. 35(8), 1798–1828 (2013)
12. Kawanobe, S., Ozaki, T.: Extraction of characteristic frequent visual patterns by
distributed representation. In: Proceedings of the 2017 31st International Con-
ference on Advanced Information Networking and Applications Workshops, pp.
525–530 (2017)
13. Kawanobe, S., Ozaki, T.: Experimental study of characterizing frequent itemsets
using representation learning. In: Proceedings of the 2018 32nd International Con-
ference on Advanced Information Networking and Applications Workshops, pp.
170–174 (2018)
14. Ozaki, T.: Evaluation measures for frequent itemsets based on distributed represen-
tations. In: Proceedings of the 2018 Sixth International Symposium on Computing
and Networking, pp. 153–159 (2018)
15. Tan, P.-N., Kumar, V., Srivastava, J.: Indirect association: mining higher order
dependencies in data. In: Proceedings of the 4th European Conference on Principles
of Data Mining and Knowledge Discovery, pp. 632–637 (2000)
16. Ras, Z.W., Dardzinska, A., Tsay, L.-S., Wasyluk, H.: Association action rules. In:
Proceedings of the 2018 IEEE International Conference on Data Mining Workshop,
pp. 283–290 (2008)
17. Kawaguchi, M., Ozaki, T.: Finding replaceable ingredients by indirect association
rules. In: Proceedings of the 78th National Convention of Information Processing
Society of Japan, vol. 1, pp. 525–526 (2016). (in Japanese)
18. Suzuki, E.: Discovering action rules that are highly achievable from massive data.
In: Proceedings of the 13th Pacific-Asia Conference on Advances in Knowledge
Discovery and Data Mining, pp. 713–722 (2009)
19. Pennington, J., Socher, R., Manning, C.D.: GloVe: global vectors for word repre-
sentation. In: Proceedings of the 2014 Conference on Empirical Methods in Natural
Language Processing, pp. 1532–1543 (2014)
20. Mikolov, T., Chen, K., Corrado, G., Dean, J.: Efficient estimation of word repre-
sentations in vector space. arXiv preprint, arXiv:1301.3781 (2013)
21. Mikolov, T., Sutskever, I., Chen, K., Corrado, G.S., Dean, J.: Distributed repre-
sentations of words and phrases and their compositionality. In: Advances in Neural
Information Processing Systems, vol. 26, pp. 3111–3119 (2013)
22. Bojanowski, P., Grave, E., Joulin, A., Mikolov, T.: Enriching word vectors with
subword information. arXiv preprint arXiv:1607.04606 (2016)
23. Arora, S., Liang, Y., Ma, T.: A simple but tough-to-beat baseline for sentence
embeddings. In: Proceedings of 5th International Conference on Learning Repre-
sentations (2017)
Estimation of Emotion Type
and Intensity in Japanese Tweets Using
Multi-task Deep Learning
Kazuki Sato1 and Tomonobu Ozaki2(B)

1
Graduate School of Integrated Basic Sciences, Nihon University, Tokyo, Japan
2
College of Humanities and Sciences, Nihon University,
3-25-40 Sakurajosui, Tokyo 156-8550, Japan
tozaki@chs.nihon-u.ac.jp
Abstract. Accurate estimation of emotions in SNS posts plays an essen-

tial role in a wide variety of real world applications such as intelligent
dialogue systems, review analysis for recommendations and so on. In this
paper, we focus on developing accurate models for estimating types of
emotions and their intensities in Japanese tweets by using multi-task
deep learning. More concretely, three deep learning models for estimat-
ing intensities of emotions were extended to be able to predict the emo-
tional types and their intensities at a time. The effectiveness of the devel-
oped models was confirmed through experiments using the database of
Japanese tweets annotated with intensity scores of four types of emotions
using best-worst scaling.
1 Introduction
Estimating emotions in SNS posts plays an essential role in a wide variety of

real world applications such as intelligent dialogue systems, reputation analysis of
products, review analysis for recommendations, prediction of national elections,
prediction of stock prices and so on. While the main target of estimation is polar-
ity and types of emotions in early stages of research, the accurate estimations
of emotional strength or intensity attract much attention in recent years. For
example, two workshops WASSA-2017 [1] and SemEval2018 [2] prepare shared
tasks to estimate the emotional intensity scores (0–1) concerning four types of
emotions (anger, fear, joy, sadness) from Twitter data. Most of these researches
on emotion mining utilizes English data resources and sophisticated dictionar-
ies on emotions such as Sentiment140 [3] and NRC Affect Intensity Lexicon [4].
On the other hand, researches on emotion analysis for Japanese tweets are not
sufficient and still in early stages. Since Japanese tweets often consist of short
and broken sentences, accurate morphological and syntactic analyses are very
difficult. In addition, electric dictionaries on emotional words in Japanese have
insufficient accuracy and range, and simple estimation by dictionary matching
cannot expect high accuracy.

https://doi.org/10.1007/978-3-030-15035-8_30
Estimation of Emotion Type and Intensity in Japanese Tweets 315
In this paper, we focus on developing accurate models for estimating emo-

tional types and intensities in Japanese Tweets. We first build a database having
8,000 of Japanese tweets annotated with intensity of four types of emotions, i.e.
joy, anger, sadness and pleasant, by using a technique called best-worst scaling
with nine annotators. We then develop a multi-task deep learning framework
[5] by extending three existing deep learning models for emotion mining to be
able to predict emotional types and intensities at a time. Using the constructed
database of Japanese tweets, the developed simultaneous estimation models are
evaluated experimentally from the aspect of estimation accuracy compared with
the case where the emotional types and intensities are predicted separately.
The rest of this paper is organized as follows. Related works are described in
Sect. 2. In Sect. 3, we explain how to construct our database of Japanese tweets
in detail. The multi-task deep learning framework for emotion estimation is
proposed in Sect. 4. After showing the experimental results in Sect. 5, we conclude
the paper and describe future direction in Sect. 6.
2 Related Work
Many methods to estimate emotional intensity have been proposed in recent
years. Himeno [6] proposed a method to estimate the emotional intensity of
tweets by combining features obtained from a model called the DeepMoji model
[7] and those obtained using multiple affect lexicons based on EmoInt [8]. The
DeepMoji model is trained through emoji predictions on a dataset of 1.2 billion
tweets with 64 common emoji labels and is a model with high performance in
tasks such as emotion analysis and sarcasm classification by transfer learning.
The workshop WASSA-2017 [1] held the first international competition on
estimating the intensity of emotion in tweets. Participants were required to esti-
mate the emotional strength of the tweet, and their accuracies were evaluated
by Pearson correlation coefficients. In the workshop, 22 teams took part in the
shared task, and commonly used features included word embeddings and sen-
tence representations learned by neural networks. The top eight systems also
made use of a substantially larger number of affect lexicons, or dictionaries,
to generate features. The best performing system was Prayas [9] with 0.747 of
Pearson correlation. The system used an ensemble of three different models:
feed-forward neural network, multi-task deep learning and sequence modeling
using Convolutional Neural Networks (CNN) [10] and Long Short-Term Memory
(LSTM) [11]. The YZU-NLP [13] proposed a combined model of Bidirectional
LSTM (BiLSTM) [12] and CNN without dictionaries to encode the global infor-
mation captured by LSTM among the most principal features extracted by CNN.
The system achieved 0.677 of correlation. As similar, the YNU-HPCC [14] with
0.671 of correlation also utilized a combined model of CNN and LSTM without
dictionaries to extract both local information within tweets and long-distance
dependency across tweets in the regression process.
Another shared task on emotional intensity estimation was held in the work-
shop SemEval2018 [2]. In the competition, SeerNet [15] won the first place by
316 K. Sato and T. Ozaki
using various state-of-the-art machine learning models, transfer learning and

ensemble learning. NTUA-SLP [16], which won the second place, adopted a
BiLSTM architecture equipped with a multi-layer self attention mechanism. It
also employed a transfer learning approach by pre-training using the dataset in
SemEval2017. The third place taker PlusEmo2Vec [17] effectively utilized exter-
nal dataset on emojis and hashtags to exploit the transferred knowledge and
used traditional machine learning algorithm such as support vector regression
and logistic regression.
3 Dataset on Japanese Tweets with Emotional Intensity

In this section, we describe how to construct a dataset of Japanese tweets with
emotional types and intensities.
Since we believe that emoticons express users’ emotions well, tweets including
emoticons are targeted in this research. First, from the collected tweets during
the period from 2014 to 2017 using Twitter streaming API, Japanese tweets
having emoticons were extracted by regular expressions. After that, 1,000 tweets
for each emotion of joy, anger, sadness and pleasant were selected by one human
expert. In addition, the same amount of 4,000 pseudo tweets were prepared by
removing emoticons from each selected tweet. As a results, we obtained 8,000
tweets in total.
In order to associate the emotional intensity with each tweet, as similar to
the case in WASSA-2017 [1], a technique called best-worst scaling was employed
with nine annotators. In the best-worst scaling, annotators repeat receiving four
tweets having the same type of emotion at a time, and selecting a tweet with
the strongest emotional intensity and one with the weakest among the four,
respectively. Based on the annotation, the emotional intensity score of each tweet
t was determined by the equation

1 #best(t) − #worst(t)
intensity(t) = +1 (1)
2 #n(t)
where #best(t) and #worst(t) are the numbers of times that tweet t was
selected as the strongest and the weakest, respectively. The constant number
#n(t) represents the total number of times the tweet t is given to the all
annotators. In this research, #n(t) = 9, i.e. the number of annotators, for all
tweets. Because the range of #best(t) and #worst(t) is [0, #n(t)], the range of
v = (#best(t) − #worst(t))/#n(t) becomes [−1, 1]. In the above equation, v
is linearly transformed by adding one and dividing by two so that the range of
intensity(t) becomes [0, 1].
The average, variance and distribution of emotional intensity of annotated
8,000 of tweets are shown in Table 1 and Fig. 1, respectively. It is observed that
original tweets having emoticons have high value of intensities compared with
those without emoticons in all four types of emotions. This tendency is noticeable
except for joy. As a result, while joy has a normal distribution in whole, sadness
and pleasant have a distribution with multimodality.
Table 1. Average and variance of emotional intensities
All tweets Without emoticons With emoticons

Average Variance Average Variance Average Variance
Joy 0.500 0.047 0.346 0.026 0.654 0.020
Anger 0.500 0.038 0.376 0.026 0.624 0.020
Sadness 0.500 0.047 0.334 0.021 0.666 0.018
Pleasant 0.500 0.049 0.327 0.020 0.674 0.017
joy anger sadness pleasant
All Tweets
Tweets without emoticons
Tweets with emoticons
Fig. 1. Distributions of emotional intensities
4 Multi-task Deep Learning for Estimating Emotional

Types and Intensities
In this research, a multi-task deep learning framework [5] for predicting emo-
tional types and intensities at a time is built by extending three existing deep
learning models shown below.
Model 1: The YZU-NLP [13] treats a tweet as a word sequence and processes it
in the order of a Bi-LSTM layer, a Convolution layer, a Max-Pooling layer,
a Dense Layer. This model is trained by the mean squared error.
Model 2: The YNU-HPCC [14] decomposes tweets into word sequences and
processes them in the order of a Convolution layer, a Max-Pooling layer,
a LSTM layer, a Dense layer. Note that, compared with YZU-NLP, the
Convolution and LSTM layers are arranged in reverse order. YNU-HPCC
employs the mean absolute error in the training.
Model 3: This model [18] combines word level representation and character level
representation. While word level inputs are processed with a Bidirectional
Gated Recurrent Unit (Bi-GRU) [19], character level inputs are processed in
the order of a Convolution layer with multiple filter sizes and a GRU layer.
These two outputs are passed into a balancer named Highway Networks [20]
to output the prediction value. While this method is originally developed
for the classification tasks using the cross entropy with L2 regularization,
we modify it for the regression tasks with mean squared error.
The architecture of our proposed multi-task learning is very simple. A module

for predicting the types of emotions using the cross entropy as loss function is
attached with each model for intensity estimation. At the same time, LSTM
layers are replaced with GRU layers. While sharing the network, the learning is
performed based on the weighted average of loss functions on intensity estimation
and classification tasks.
5 Experiments
To assess the effectiveness of the proposed framework, we conducted experi-

ments using the database of Japanese tweets annotated with intensity of four
types of emotions. The 8,000 of tweets were divided into three: 5,120 for train-
ing, 1,280 for validation and 1,600 for test. The stratified division is applied
so that each emotional type has 1,280 in training, 320 in validation and 400
in tests, respectively. As baseline methods, by removing irrelevant module from
our proposal, we prepared models for estimating only emotional intensities and
those for predicting only types of emotions. The hyperparameters in each model
were determined to maximize the sum of Pearson correlation and classification
accuracy in the validation data by using Bayesian optimization. In the learning,
the adam [21] was employed as an optimization method and the epoch number
is set to 30. If the loss value does not improve in 5 epochs, we judge learning has
converged and abort the learning.
The obtained hyperparameters and estimation results are shown in Tables 2
and 3, respectively. The estimation results of emotional intensities are also
depicted in Fig. 2. The best models in multi-task learning with respect to the
evaluation measures are summarized in Table 4. As similar to the case in the
workshop WASSA-2017 [1], Pearson correlation is employed to evaluate the esti-
mation accuracy of intensities.
As shown in the Table 3, since most results of precision, recall and f-measure
on classifying the types of emotions by the multi-task learning exceed 0.9, we can
conclude that the classification performance is very high. Regarding the classifi-
cation tasks, it is observed in the Table 4 that Model 2 and Model 3 are superior
to Model 1 as a whole. While Model 3 is the best for precision, Model 2 is the best
for recall. These two have similar performance with respect to f-measure.
Table 2. Hyper parameters
Multi-task Classification Regression

Model 1
Dimension of word vectors 200 100 100
BiGRU hidden state size 500 100 500
CNN filter size 4 2 2
CNN feature maps 16 32 64
Pooling filter size 3 4 2
Dense layer size 16 64 64
Dropout rate 0.1 0.5 0.0
Batch size 16 16 16
Model 2
CNN filter size 3 3 3
CNN feature maps 32 64 16
Pooling filter size 2 3 2
Dropout rate 0.0 0.4 0.0
GRU hidden state size 200 200 300
Batch size 64 32 16
Model 3
Dimension of character vectors 50 50 100
BiGRU (word) hidden state size 200 100 300
CNN feature maps (filter size 2) 16 16 64
GRU (character) hidden state size 200 500 300
Batch size 64 32 16
Regardless of types of emotions, Model 3 is the best on the estimation per-

formance of emotional intensities, and Model 2 always outperforms Model 1.
We can observe in the Fig. 2 that Model 1 has narrow prediction range and it
fails to estimate large and small intensity scores accurately. On the other hand,
although the residuals are not small, it is recognized that Model 3 have a posi-
tive correlation between the predicted and actual values. From these results, we
can conclude that Model 3 has high performance as a whole.
Model 1 Model 2 Model 3
Fig. 2. Intensity scores (x-axis: actual values, y-axis: predicted values)
Table 3. Experimental results: estimation accuracy
Multi-task learning Separate learning

Precision Recall F-measure Correlation Precision Recall F-measure Correlation
All
Model 1 0.907 0.907 0.906 0.234 0.920 0.915 0.917 0.622
Model 2 0.920 0.923 0.922 0.610 0.920 0.922 0.921 0.603
Model 3 0.921 0.922 0.922 0.776 0.934 0.936 0.935 0.771
Joy
Model 1 0.922 0.869 0.884 0.161 0.923 0.901 0.903 0.639
Model 2 0.922 0.933 0.926 0.628 0.920 0.933 0.922 0.609
Model 3 0.897 0.939 0.915 0.795 0.924 0.950 0.934 0.792
Anger
Model 1 0.944 0.926 0.924 0.415 0.949 0.918 0.924 0.600
Model 2 0.928 0.915 0.919 0.589 0.933 0.928 0.926 0.600
Model 3 0.946 0.924 0.933 0.686 0.938 0.959 0.945 0.701
Sadness
Model 1 0.839 0.912 0.858 0.382 0.880 0.924 0.892 0.608
Model 2 0.888 0.921 0.902 0.621 0.925 0.882 0.897 0.613
Model 3 0.894 0.920 0.904 0.770 0.924 0.921 0.918 0.757
Pleasant
Model 1 0.922 0.919 0.911 0.027 0.930 0.915 0.914 0.647
Model 2 0.944 0.923 0.932 0.605 0.901 0.944 0.918 0.599
Model 3 0.948 0.906 0.924 0.836 0.953 0.915 0.931 0.823
Table 4. Best models in multi-task learning
Precision Recall F-measure Correlation

All Model 3 Model 2 Model 2, Model 3 Model 3
Joy Model 1, Model 2 Model 3 Model 2 Model 3
Anger Model 3 Model 1 Model 3 Model 3
Sadness Model 3 Model 2 Model 3 Model 3
Pleasant Model 3 Model 2 Model 2 Model 3
In the Table 2, the obtained hyperparameters are greatly different between

the multi-task learning and separate learning. Since the same can be observed
for the classification and regression tasks in separate learning, it can be seen
that the suitable hyperparameters are largely different for these two tasks. This
results show that it is not easy to determine the appropriate hyperparameters
for the multi-task learning.
The positive effect on multi-task learning can be observed in the estimation
tasks of intensity in Model 2 and Model 3 except for the anger emotion. However,
the positive effect cannot be recognized with respect to the classification tasks.
The performance is improved in only six cases listed below: f-measure for joy by
Model 2, f-measure for sadness by Model 2, f-measure for pleasant by Model 2,
precision for anger by Model 3, recall for pleasant by Model 1 and precision
for pleasant by Model 2. These results may be caused by the consideration of
classification and regression tasks to the same extent. We need to consider the
loss function for learning and the criteria for hyperparameter tuning to improve
the total performance.
6 Conclusion
In this paper, to achieve accurate estimations of emotional types and intensi-
ties of Japanese tweets, we constructed a database of Japanese tweets anno-
tated for joy, anger, sadness and pleasant intensities using the best-worst scaling
method. In addition, we proposed a multi-task deep learning framework based on
three existing models. Through the experiments, we confirmed that the proposed
framework can improve the estimation accuracy on emotional intensities.
As future work, it is necessary to evaluate the proposed framework using large
scale datasets with more emotional types. It is also necessary to develop certain
mechanisms for balancing the classification and regression tasks. We also plan
to develop a model which combines the unsupervised representation learning for
tweets and the ensemble learning with multiple estimators.
Acknowledgements. This work was partially supported by JSPS KAKENHI Grant

Number 17K00315.
References
1. Mohammad, S.M., Bravo-Marquez, F.: WASSA-2017 shared task on emotion inten-
sity. In: Proceedings of the 8th Workshop on Computational Approaches to Sub-
jectivity, Sentiment and Social Media Analysis (2017)
2. Mohammad, S.M., Bravo-Marquez, F., Salameh, M., Kiritchenko, S.: SemEval-
2018 task 1: affect in Tweets. In: Proceedings of the 12th International Workshop
on Semantic Evaluation (2018)
3. Go, A., Bhayani, R., Huang, L.: Twitter sentiment classification using distant
supervision. In: Final Projects from CS224N for Spring 2008/2009 at The Stanford
Natural Language Processing Group (2009)
4. Mohammad, S.M.: Word affect intensities. arXiv preprint arXiv:1704.08798 (2017)
5. Keras Documentation: Multi-input and multi-output models. https://keras.io/

getting-started/functional-api-guide/#multi-input-and-multi-output-models
6. Himeno, S., Aono, M.: Estimation of Twitter emotion intensity by using transfer
learning of feature tensors and emotion lexicon. IEICE Technical report DE2018-11
(2018). (in Japanese)
7. Felbo, B., Mislove, A., Søgaard, A., Rahwan, I., Lehmann, S.: Using millions of
emoji occurrences to learn any-domain representations for detecting sentiment,
emotion and sarcasm. In: Proceedings of the 2017 Conference on Empirical Meth-
ods in Natural Language Processing, pp. 1615–1625 (2017)
8. Duppada, V., Hiray, S.: Seernet at EmoInt-2017: tweet emotion intensity estimator.
In: Proceedings of the 8th Workshop on Computational Approaches to Subjectivity,
Sentiment and Social Media Analysis, pp. 205–211 (2017)
9. Goel, P., Kulshreshtha, D., Jain, P., Shukla, K.K.: Prayas at EmoInt 2017: an
ensemble of deep neural architectures for emotion intensity prediction in tweets.
In: Proceedings of the 8th Workshop on Computational Approaches to Subjectivity,
Sentiment and Social Media Analysis, pp. 58–65 (2017)
10. Kim, Y.: Convolutional neural networks for sentence classification. In: Proceedings
of the 2014 Conference on Empirical Methods in Natural Language Processing, pp.
1746–1751 (2014)
11. Hochreiter, S., Schmidhuber, J.: Long short-term memory. Neural Comput. 9(8),
1735–1780 (1997)
12. Schuster, M., Paliwal, K.K.: Bidirectional recurrent neural networks. IEEE Trans.
Sig. Process. 45, 2673–2681 (1997)
13. He, Y., Yu, L.-C., Lai, K.R., Liu, W.: YZU-NLP at EmoInt-2017: determining
emotion intensity using a bi-directional LSTM-CNN model. In: Proceedings of the
8th Workshop on Computational Approaches to Subjectivity, Sentiment and Social
Media Analysis, pp. 238–242 (2017)
14. Zhang, Y., Yuan, H., Wang, J., Zhang, X.: YNU-HPCC at EmoInt-2017: using a
CNN-LSTM model for sentiment intensity prediction. In: Proceedings of the 8th
Workshop on Computational Approaches to Subjectivity, Sentiment and Social
Media Analysis, pp. 200–204 (2017)
15. Duppada, V., Jain, R., Hiray, S.: SeerNet at SemEval-2018 task 1: domain adap-
tation for affect in tweets. In: Proceedings of the 12th International Workshop on
Semantic Evaluation, pp. 18–23 (2018)
16. Baziotis, C., Nikolaos, A., Chronopoulou, A., Kolovou, A., Paraskevopoulos, G.,
Ellinas, N., Narayanan, S., Potamianos, A.: NTUA-SLP at SemEval-2018 task 1:
predicting affective content in tweets with deep attentive RNNs and transfer learn-
ing. In: Proceedings of the 12th International Workshop on Semantic Evaluation,
pp. 245–255 (2018)
17. Park, J.H., Xu, P., Fung, P.: PlusEmo2Vec at SemEval-2018 task 1: exploiting
emotion knowledge from emoji and #hashtags. In: Proceedings of the 12th Inter-
national Workshop on Semantic Evaluation, pp. 264–272 (2018)
18. Liang, D., Xu, W., Zhao, Y.: Combining word-level and character-level representa-
tions for relation classification of informal text. In: Proceedings of the 2nd Work-
shop on Representation Learning for NLP, pp. 43–47 (2017)
19. Cho, K., Merrienboer, B., Gulcehre, C., Bougares, F., Schwenk, H., Bengio, Y.:
Learning phrase representations using RNN encoder-decoder for statistical machine
translation. In: Proceedings of the 2014 Conference on Empirical Methods in Nat-
ural Language Processing, pp. 1724–1734 (2014)
20. Srivastava, R.K., Greff, K., Schmidhuber, J.: Training very deep networks. In:
Proceedings of the 28th International Conference on Neural Information Processing
Systems, vol. 2, pp. 2377–2385 (2015)
21. Kingma, D., Ba, J.: Adam: a method for stochastic optimization. arXiv preprint,
arXiv:1412.6980 (2014)
A Method for Extracting Influential People
for the Improvement of Contents
Hayato Tsukiji1 and Kosuke Takano2(&)

1
Course of Information and Computer Sciences,
Graduate School of Engineering, Kanagawa Institute of Technology,
Atsugi, Kanagawa, Japan
s1785005@cco.kanagawa-it.ac.jp
2
Department of Information and Computer Sciences,
Faculty of Information Technology, Kanagawa Institute of Technology,
Atsugi, Kanagawa, Japan
takano@ic.kanagawa-it.ac.jp
Abstract. It is very important to find influential users who have useful infor-
mation as well as extract helpful knowledge that affect the efficiency of user’s
intellectual work; however, it is difficult to evaluate how such helpful comments
and conversations from influential users can affect works of other users. In this
study, we propose a method for extracting user comments that influenced other
users to improve their contents such as a presentation slide by analyzing cor-
relation between user’s comments and improved parts of contents. Our method
allows us to make actual evaluation to influential users for some specific users.
In the experiment, we evaluate the feasibility of the proposed method using the
actual communication history that is collected from a communication service,
Slack.
1 Introduction
Information sharing services and collaborative developing services are widely used for
sharing the knowledge and information on intellectual and productive activities such as
software development and academic research. Using these services, users are doing
various activities such as discussion, work instruction, content share, and schedule
management for the purpose of knowledge and information sharing, and expertise
knowledge that users have is accumulated in various forms such as conversation
stream, document, presentation slide, and source code.
In this study, we focus on ‘influential users’ who often give advice and provide
useful information to some users in such activities of information sharing. By lever-
aging knowledge and information of influential users, it is deemed that the cost of
information sharing with other users can be reduced and the motivation to effectively
use information for some specific purpose can be improved. In our previous study, we
have proposed a method for extracting users who have influential roles, such as leader,
moderator, supporter, and facilitator, that are extracted from the technical communi-
cation history for extracting influential users for other users in a communication group
[1]. In this method, the frequency of messages and the number of users who are

https://doi.org/10.1007/978-3-030-15035-8_31
A Method for Extracting Influential People for the Improvement of Contents 325
communicated with are considered for extracting user’s roles, but how the knowledge
and information of the influential users who have special roles can affect other users are
not considered.
It is deemed that the influential users have the characteristics of ‘having useful
expertise knowledge’, ‘having reliable information’, ‘having excellent capability of
communication and information sharing’, ‘having instructional skill and problem-
solving ability’, and so on. Therefore, when a user is seeking the desired information,
from the above reasons, there is a possibility that the user can efficiently obtain useful
and reliable information by extracting information that can be derived from users who
already build reliable relationships with others, and can continue the meaningful
communication with them. Thus, it is very important to find influential users who have
such useful information as well as extract helpful knowledge from accumulated
information [2–4], which affect the efficiency of user’s intellectual work; however, it is
difficult to evaluate how such helpful comments and conversations from influential
users can affect works of other users, even if they were actually effective for improving
the efficiency of user’s work and motivating users to take action according to their
information.
In this study, for evaluating how the knowledge and information from influential
users are used by other users on an online community of technical information sharing,
we propose a method for extracting user comments that influenced other users to
improve their contents such as a presentation slide by analyzing correlation between
user’s comments and improved parts of contents. Our method allows us to make actual
evaluation to influential users for some specific users.
In the experiment, we evaluate the feasibility of the proposed method using actual
data set up by collecting presentation slides which are updated every week for research
discussion in a laboratory and user’s comments through the discussion for improving
their slides and researches.
2 Proposed Method
Figure 1 shows an overview of the proposed method. First, our method calculates
correlation between user’s comments and difference of contents updated by other users.
If correlation value is high, we can assume that the content might be improved based on
the user’s comments. Based on this assumption, our method extracts the user comments
that influenced some specific users to improve their contents and regards a user who
have such useful comments as an influential user.
Tables 1 and 2 show an example of user comments and an example difference of
sentences between old slide and new slide, respectively. As shown in Table 1, com-
ments from each user consists of several sentences for the purpose of improving one
presentation slide. In addition, Table 2 shows that there are five updated parts of
sentences in a presentation slide based on user’s comments.
326 H. Tsukiji and K. Takano
User
Extract the difference of content and content
updated in a certain period
Content C1 Content C1
Content C2 Update Difference
Content C’1 Most influential comment

…
Comment msg3
Comment msg1
User C
User A Comment msg6
User B Comment msg2
User D
User C Comment msg3 Comment msg4
User D
…
…
User B
Get comments from other users in a certain period By calculating correlation between comment
and difference of updated content, influential
comments are extracted.
Fig. 1. Overview of proposed method
Table 1. Example of comments of users

User Comments
User1 • Voice is small
• Some spaces between lines are narrow
• Sentences in background are easy to understand
• ‘A leaner has a high interest’ looks wrong assertion
• The meaning of ‘subsequent life’ is not clear. Please use an alternative expression like ‘situations
that we speak English in the future’
• Did you explain the idea of your method?
• It is useful if we have phonetic symbols. Please look for databases of phonetic symbols that might
be opened online to the public
User2 • Did you decide the environment of experiments?
User3 • What is ‘expression that matches learner’?
• It might be time that you should mention about ‘Experimental Method’
• Please write a note of oral discussion through the seminar in a top column for comments!
Table 2. Example of difference of sentences in slides

Id Difference of sentences
d1 An experimenter selects according to the difficulty of the question
d2 As he/she speaks English
d3 This study proposes a system that preferentially provides such questions of English sentences
d4 In the case that the system randomly provides questions of English sentences stored in the database
d5 As the result of study where a learner solve questions that is selected based on the shared content
Set of contents C Set of comments MSG for

User u1
set of contents
C1 C2 Ci C
… …
MSG(C1)
Δt1 Δt2 Δt3 … Δtk MSG(C2)
…
Ci Ci Ci … Ci
Other users
u2 un
MSG(Ci)
…
ΔC1 ΔC2 … ΔCl-1
MSG(Cn)
Difference in a period Δt1 Comments for content

Calculate correla on in a period Δt1
Sentence d1
MSG2(Ci, Δt1)
Sentence d2
MSG3(Ci, Δt1)
…
…
Sentence dm
MSGn(Ci, Δt1)
Rank users who is the most influen al for
Extract influen al comments for other users by calcula ng similarity score user u1 based on the sum total of similarity
between difference ΔC and comment MSG(Ci) score in a certain period t
MSG2(Ci, Δt1) MSG5(Ci, Δt1)

User u6 Most influential
MSG6(Ci, Δt1) MSG6(Ci, Δt1) User u1
User u3 user for u1
ΔC1 … ΔCl-1
Fig. 2. Execution steps of proposed method
The proposed method is executed in the following steps (Fig. 2).

Step-1: A set of contents C that consist of n contents that a target user ux has is
collected.
Step-2: For each content ci in C, differences Dc1 ; Dc2 ; . . .; Dcl1 within a constant
period Dtk ðk ¼ 1; 2; . . .; l 1Þ are obtained. Then, sentences d1 ; d2 ; . . .; dm
are extracted from each Dcr .
Step-3: A set of user’s comments MSGy(ci) that are given by other user uy ðy 6¼ xÞ
for a content ci is extracted. In addition, a set of user’s comments
MSGy ðci ; Dtk Þ, which is a subset of MSGy for Dtk , is extracted.
Step-4: In Steps-2 and 3, similarity scores are calculated between MSGy(ci, Δtk) and
each difference of sentence dj. Here, MSGy(ci, Δtk) and dj are vectorized
using a term weight such as TFIDF, and a similarity score is calculated using
a cosine measure and so on in the vector space model. From MSGy(ci, Δtk)
given by other user uy, the comments with higher correlation scores are
extracted as influential user comments that affect the improvement of
content.
In addition, by applying the calculation in Step-4 for all set of comments in a period
t, and the total sum of the similarity scores are calculated for each uy (y 6¼ x). Users are
ranked based on the total sum of similarity scores, and the users ranked in the top-p are
extracted as users who are influential for a user ux to improve the content.
3 Experiments
In the experiment, we used the actual data set by collecting presentation slides which
are updated every week for research discussion in a laboratory and user’s comments
through the discussion for improving their slides and researches. Table 3 shows the
number of experimenters. Three users User1, User2, and User3 only gave their com-
ments and eight users User4 to User11 updated their slides as well as gave comments to
other users. In addition, Tables 4 and 5 show the number of difference of sentences in
slides and the number of comments received from other users, respectively.
From eight users, we focus on the result of User8 and User10 whose frequency of
updating slide is high. Tables 6, 7, and 8 show user comments for User8, difference of
sentences in slide of User8, and correlation values between user comments and dif-
ference of sentences by User8, respectively. Similarly, Tables 9, 10, and 11 show the
result of User10. From Table 8, we can find that correlation values for User1, User2,
and User10 are high, and this means User8 improves the slides based on these three
users. From these results, it is deemed that comments from User1, User2, and User 10
are influential for User8. Similarly, from Tables 9, 10, and 11, we can confirm that
comments from User1, User3, and User6 are influential for User10.
Table 3. Number of experimenters

Number
User who gave comments 3
User who gave comments and updated a slide 8
Total 11
Table 4. Number of difference of sentences in slides

User Difference of sentences
User4 10
User5 8
User6 9
User7 4
User8 10
User9 10
User10 9
User11 10
Table 5. Number of comments received from other users

Seminar ID User4 User5 User6 User7 User8 User9 User10 User11
1 5 4 4 0 8 7 5 7
2 4 8 6 6 6 0 7 5
3 6 0 5 0 6 6 4 5
4 6 0 5 0 5 7 7 8
5 3 4 5 7 3 4 5 6
6 4 5 0 4 3 4 4 6
7 4 5 6 6 6 0 4 4
8 3 4 7 0 4 3 4 3
9 2 4 4 0 2 3 3 3
10 0 3 5 3 3 0 0 0
11 0 0 0 0 0 0 0 0
Total 37 37 47 26 46 34 43 47
Table 6. Comments for User8

User Comment
User1 • Slide5: Please explain the meaning of “speech that is easy for a learner to hear”
• Slide7: Where do you explain how your method calculates “easiness of hearing”?
• Slide11: If correct rate = “easiness of hearing”, please use same expression
• Please describe first “we can expect that continuity is improved by starting from
speech that is easy for a learner to hear”. (If not mentioned)
• Please add random and difficulty in order as a baseline
• Make a strategy for experiment scenario, environment, number of experimenters
• Do enough test of system usability before experiments
Usre2 • How degree the skill of experimenters is considered?
• Number of people who studied abroad are defined in advance, or they are
randomly selected
User4 • p12 the meaning of false arrow is unclear
• p25 questionnaire for learning motivation is taken every day? - > In p31, I can
understand or cannot …
User10 • It is difficult for people who listened first time to understand the meaning of
“keeping motivation”
Table 7. Difference of sentences in slide of User8

Id Sentence
4 A lot of pronunciation that is easy for a learner to hear are included
67 Estimate from learning history, and a leaner from speech of estimated area
95 A learner with English communication ability
101 Improvement of learning motivation*
213 According to following steps, provide questions to a learner. Set the number of
questions by region
Table 8. Correlation values between user comments and difference of sentences by User8
Id User1 User2 User4 User10
4 0.189 0.000 0.000 0.000
67 0.157 0.040 0.133 0.135
95 0.089 0.079 0.067 0.135
101 0.058 0.000 0.086 0.174
213 0.121 0.160 0.135 0.091
Table 9. Comments for User10

User Comment
User1 • Example is needed in many cases
• It is difficult to understand precisely, when we listen to unknown idea, if you explain
only procedure
• Example of “questions that match with”
User2 • It is better to include example for why that approach is promising (ex: we can
expect the learning effect by focusing on English words in a research paper than a
standard way, if the purpose is to practice oral presentation in a conference)
User3 • Use itemized style than newline for “Background”
• “Major implementations” There are two, although it says “following three”
• “Research purpose” How do you evaluate “keeping high motivation of learning”
• Show example of “expression that matches better with learner”
User6 • It says “a learner had question that is matched with”, but it might change according
to genre of question?
User7 • Difference of newline and paragraph is unclear in two slides
User8 • Comment
• It’s better to insert example figures -> Slides 2 to 5
• Please clearly explain advantages of proposed system using some figures
User9 • Since contents a learner has interests in are written in Japanese, might it be good to
translate it into English for question?
Table 10. Difference of sentences in slide of User10

Id Text
22 Match
23 Example of English sentences
29 Example of English sentences that don’t match with
Table 11. Correlation values between user comments and difference of sentences by User10
Id User1 User2 User3 User6 User7 User8 User9
22 0.177 0.000 0.139 0.267 0.000 0.000 0.000
23 0.306 0.091 0.240 0.154 0.000 0.183 0.115
29 0.395 0.071 0.248 0.239 0.169 0.141 0.089
In addition, Tables 12, 13, 14, and 15 shows the ranking results of total sum of
correlation values for four users: User4, User8, User10, and User11, respectively.
In Table 8, User2, User1, and User8 are ranked higher. This is because User4 has
improved slides according to their comments, and we can see that User2, User1, and
User8 are influential users for User4. Similarly, it is deemed that User2 and User1 are
influential for User4, User2, User3, and User1 are influential for User10, User4 and
User2 are influential for User 11.
From these result, we can confirm that proposed method can extract influential
users based on correlation between user’s comments and improved parts of contents.
Table 12. Ranking result for User4

User Score Count
User2 46.813 9
User1 44.349 5
User8 23.890 7
User3 5.002 3
User10 4.749 2

User Score Count
User2 131.766 11
User1 95.542 6
User4 92.626 10
User10 33.148 6
User3 20.129 5

User Score Count
User2 43.950 10
User3 28.626 9
User1 26.977 7
User4 17.255 8
User8 13.360 7

User Score Count
User4 46.405 8
User2 42.113 9
User1 25.758 5
User3 23.503 6
User8 17.452 7
4 Conclusion
This paper has presented a method for extracting user comments that influenced other
users to improve their contents by analyzing correlation between user’s comments and
improved parts of contents. In the experiment, we evaluated the feasibility of the
proposed method using the actual communication history that is collected from a
communication service, Slack, and confirmed that our method can extract influential
users by calculating correlation values between user’s comments and improved parts of
contents.
In our future work, we will evaluate the effectiveness of our method using larger
size of communication history in longer period. In addition, we will develop a rec-
ommendation service to provide useful knowledge and information of influential users
by applying our method to recommendation algorithm.
References
1. Hayato, T., Kosuke, T.: An estimation method of user roles based on online technical
communication activities. In: Proceedings of the Workshops of the 32nd International
Conference on Advanced Information Networking and Applications (WAINA-2018),
pp. 181–186 (2018)
2. Khan, H.U., Daud, A., Malik, T.A.: MIIB a metric to identify top influential bloggers in a
community. PLoS ONE 10(9), e0138359 (2015)
3. Akritidis, L., Katsaros, D., Bozanis, P.: Identifying the productive and influential bloggers in
a community. IEEE Trans. Syst. Man Cybern. Part C Appl. Rev. 41(5), 759–764 (2011)
4. Ishfaq, U., Khan, H.U., Khalid, I.: Identifying the influential bloggers a modular approach
based on sentiment analysis. J. Web Eng. 16(5&6), 505–523 (2017)
5. Resnick, P., Iacovou, N., Suchak, M., Bergstrom, P, Riedl, J.: GroupLens an open
architecture for collaborative filtering of netnews. In: CSCW 1994, Proceedings of the 1994
ACM Conference on Computer Supported Cooperative Work, pp. 175–186 (1994)
6. Truong, K.Q., Ishikawa, F., Honiden, S.: Improving accuracy of recommender system by
item clustering. IEICE Trans. Inf. Syst. 90(9), 1363–1373 (2007)
7. Tan, L.K.W., Na, J.C., Theng, Y.L.: Influence detection between blog posts through blog
features, content analysis, and community identity. Online Inf. Rev. 35(3), 425–442 (2011)
8. Zhang, S., Zhang, S., Yen, N.Y., Zhu, G.: The recommendation system of micro-blog topic
based on user clustering. Mobile Netw. Appl. 22(2), 228–239 (2017)
9. Li, W., Ye, Z., Xin, M., Jin, Q.: Social recommendation based on trust and influence in SNS
environments. Multimed. Tools Appl. 76(9), 11585–11602 (2017)
10. Seo, Y.-D., Kim, Y.-G., Lee, E., Baik, D.-K.: Personalized recommender system based on
friendship strength in social network services. Expert Syst. Appl. 69(1), 135–148 (2017)
11. Basher, A.R.M.A., Fung, B.C.M.: Analyzing topics and authors in chat logs for crime
investigation. Knowl. Inf. Syst. 39(2), 351–381 (2014)
12. Yasin, M., Qureshi, J.A., Kausar, F., Kim, J.: A granular approach for user-centric network
analysis to identify digital evidence. Peer Peer Netw. Appl. 8(5), 911–924 (2015)
13. Wang, T., Huang, Z., Gan, C.: On mining latent topics from healthcare chat logs. J. Biomed.
Inform. 61, 247–259 (2016)
Sentiment Analysis of Arabic
and English Tweets
Mohamed K. Elhadad, Kin Fun Li(&), and Fayez Gebali
University of Victoria, Victoria, Canada

{melhaddad,kinli,fayez}@uvic.ca
Abstract. Due to the continuous and rapid growth of daily posted data on the
social media sites in many different languages, the automated classification of
this huge amount of data has become one of the most important tasks for
handling, managing, and organizing this huge amount of textual data. There
exist many examples of social media sites, but Twitter is considered to be one of
the most popular and commonly used, as users are able to communicate with
each other, share their opinions, and express their emotions (sentiments) in the
form of convenient short blogs using less than 140 words. Accordingly, many
companies and organizations may analyze these sentiments in order to evaluate
the users’ thoughts, and determine their polarity from the content of the text. For
this process, natural language processing techniques, statistics, or machine
learning algorithms are being used to identify and extract the sentiment of the
text. In practice, many data mining techniques and algorithms are being applied
to observe patterns and correlation among that huge amount of data. This paper
proposes an efficient approach in handling Tweets, in both Arabic and English
languages, with different processing techniques applied. This approach is based
on using the Vector Space Model (VSM) to represent text documents and
Tweets, and the Term Frequency Inverse Document Frequency (TFIDF) in a
term weighting process to generate the feature vector for classification process.
The proposed approach has been evaluated using several experiments with
different classifiers on five datasets: Decision trees, Naive-Bayes, kNN, Logistic
Regression, Perceptron, and Multilayer Perceptron. The experimental results
reveal the effectiveness of our proposed approach when comparing classification
results with the published work in [1–3].
1 Introduction
In the past twenty years, social network sites have made the entire world connected
together. Using these networks, users are able to share many things, such as views,
feelings, experiences, advice, etc. These electronic Word of Mouth (eWOM) statements
expressed on the web could be beneficial and helpful to others [4, 5]. There are many
social network sites but Twitter is one of the most popular and commonly used social
network sites, where users can easily communicate with each other or share their
opinions in the form of short blogs less than 140 words. The huge amount of posted
text on Twitter is increasing in an astronomical pace. This rapid growth in data makes it
difficult to handle especially in a real-time environment with existing databases and
techniques.
https://doi.org/10.1007/978-3-030-15035-8_32
Sentiment Analysis of Arabic and English Tweets 335
Sentiment analysis has become one of the most widely researched topics in dif-
ferent aspects. The word sentiment refers to the feeling, thought, or opinion, of a person
across some particular domains. Hence, sentiment analysis is also referred to as opinion
mining. The huge amount of Tweets posted daily makes Twitter data one of the most
essential data volumes that deal with business, industry or social issues. In fact, as the
posted twitter data satisfies the properties that it is huge in size, comes from hetero-
geneous sources, is in both structured and unstructured forms, has high speed in data
generation, and is very inconsistent, so it could be considered as big data [6]. The main
goal behind the analysis of sentiments, is to evaluate these thoughts, determine its
polarity using NLP (Natural Language Processing), tagging, statistics, or machine
learning algorithms to identify and extract the sentiment content of a text.
Technically, data mining’s main goal is to observe correlations or patterns among a
huge amount of data in large relational databases. Nowadays it is not necessary to turn
to large structured databases to have large volume of data analyzed [7]. The traditional
way in acquiring knowledge from data relies on the manual analysis and interpretation
carried out by human experts who are familiar with the data domain. They serve as an
interface between the data and the users. Unfortunately, this manual approach is slow,
expensive, and highly subjective, as well as being dependent on the experience of the
experts, and it has become impracticable due to the dramatically increased volume of
data on the web [8].
In this work, the automated handling of Tweets, in both Arabic and English lan-
guages, was investigated and different processing techniques were applied on five
different datasets of various sizes in both languages. Section 2 shows the related work.
Section 3 presents the proposed multi-stage model for sentiment analysis, and shows
the details of each stage. Experimental setup and discussion are given in Sect. 4.
Section 5 summarizes the overall conclusion and expected directions for future work.
2 Related Work
Sentiment Analysis is the area of web data mining which deals with evaluating the
feelings or thoughts of people across some particular domains, and to determine the
polarity and to what degree [9]. In the past few years, the field of detecting and
evaluating the feelings of users or customers in many domains has been extensively
investigated. This is due to the growth of data which exists online over the internet,
especially those that deal with sentiment analysis such as describing people’s point of
view, experiences, preferences, and thoughts.
Neethu and Rajasree [10] and Jindal et al. [11] gave detailed and deep reviews on
machine learning approaches, documents representation techniques, datasets, and
techniques used for text-documents classification. It is clearly indicated from the
reviews that the Vector-Space-Model (VSM) is the most commonly used method for
document and text classification, with each document being represented as a compact
set of words. Ravi et al. [9] introduced a survey that covers the different types of
sentiment analysis, datasets used, natural language processing techniques used, and
their applications, as well as explanation and categorization of some sentiment analysis
algorithms and their originating references. Also, Yang et al. [12] introduced the
336 M. K. Elhadad et al.
commonly used sentiment analysis techniques, including Support Vector Machine,

Naive Bayes, Maximum Entropy, and Artificial Neural Network. Furthermore, they
discussed these techniques’ performance assessment and difficulties.
Pang et al. [13] were the first to apply machine learning for sentiment mining on
movie reviews corpus. Unigram and bag of words were utilized to obtain features. The
ratio of accuracy differs depending on their applications. Das et al. [14], meanwhile,
made an attempt for making financial decisions, such as predicting the potential prices
of a company’s stock. Kim et al. [15] found that if features are highly dependent on
other variables, Naive Bayes works relatively well. This may be counterintuitive
because Naive Bayes uses features that are independent of each other. Niu et al. [16]
used a new model with well thought out approaches used for feature selection, weight
computation, and classification. This was based on the Bayesian algorithm in which the
weights of the classifier are modified by using both representative feature and unique
feature. ‘Representative feature’ deals with the information that represents a class and
‘unique feature’ deals with the information that aids in distinguishing classes from each
other. Probabilities were calculated using this method on each classification. This
helped improve the effectiveness of Bayesian algorithm.
Barbosa et al. [17] created a 2-step automatic sentiment analysis method for
classifying Tweets. A noisy training set was used to reduce the effort required for
labeling in developing classifiers. They initially classified Tweets into subjective and
objective Tweets. Next, subjective Tweets were classified as positive and negative
Tweets. Celikyilmaz et al. [18] created a pronunciation based word clustering method
that could normalize noisy Tweets. In pronunciation based word clustering, words with
a similar pronunciation are grouped together and labeled as common tokens. Addi-
tionally, they also assigned similar tokens for numbers, html links, user identifiers, and
target organization names for normalization. Finally, after normalization, they used
probabilistic models to identify polarity lexicons for features. Classification using the
BoosTexter classifier with these polarity lexicons as features was performed and a
reduced error rate was obtained.
Nabil et al. [2] introduced ASTD, an Arabic social sentiment analysis dataset
gathered from Twitter. The 10,000 Tweets in this dataset are classified as objective,
subjective positive, subjective negative, and subjective mixed. They presented the
properties and the statistics of the dataset and run experiments using standard parti-
tioning of the dataset. Guellil et al. [19] present an approach to automatically classify
sentiments of Arabic text which relies on Latin letters, numerals and punctuation rather
than Arabic letters (Arabizi). They translated it into Arabic, then applied different
machine learning algorithms, in order to make the automated classification process of
the sentiment, using Support Vectors Machine (SVM) and Naive Bayes (NB). The
obtained results demonstrated the NB algorithm outperforms all others. The highest
achieved F1-score is up to 78% and 76%, for manually and automatically transliterated
dataset, respectively.
Baly et al. in [3] conducted an analytical study to observe challenges to perform

opinion mining in a language which is rich in morphology such as the Arabic language.
They applied different models on the same dataset in [2]. The obtained accuracy results
were 49.5% when applying the Support Vector Machine (SVM), and 58.5% when
applying the Recursive Neural Tensor Networks (RNTN). Heikal et al. [4] applied
previously unused and different deep learning models to improve sentiment analysis
accuracy of Arabic text. They used an ensemble model, combining Convolutional
Neural Network (CNN) and Long Short-Term Memory (LSTM) models, to predict the
sentiment of Arabic Tweets. The obtained results after applying their model on the
Arabic Sentiment Tweets Dataset (ASTD), achieves an accuracy of 64.3%, 64.75%,
65.05%, with CNN, LSTM, and the ensemble models, respectively [2].
Bhavitha et al. in [20] compared different classifiers for sentiment analysis. This
was done on Twitter posts related to electronic products such as mobiles and laptops.
They utilized various classifiers including Naive Bayes, Support Vector Machine,
Maximum Entropy, and Ensemble for text classification and compared their accuracy,
precision, and recall. A dataset of 1200 twitter posts divided equally into positive and
negative sentiment classes was used as training (1000) and test (200) sets. Results
concluded that the Naive Bayes classifier has the highest precision, but lacks accuracy
and recall. Rane et al. [1] worked on a dataset comprising of Tweets for 6 major US
Airlines and performed a multi-class sentiment analysis. This approach starts with pre-
processing techniques used to clean the Tweets and then represents these Tweets as
vectors using a deep learning concept (Doc2vec) to do a phrase-level analysis. They
applied different classification techniques, and uses the precision, recall, and F-measure
to evaluate the performance of the classifiers.
Dashtipour et al. [21] present a state-of-the-art review on multilingual sentiment
analysis showing independent comparison of techniques and features used. Balahur [22]
dealt with the problem of sentiment detection in three different languages—French,
German and Spanish—using three distinct Machine Translation (MT) systems: Bing,
Google and Moses. They claimed the superiority of their system, but did not taking into
consideration that the translation process is not accurate enough and takes time.
In summary, there are many existing works in the field of sentiment analysis of
Tweets applied to different languages. In this work, another approach is proposed to
deal with English and Arabic languages with the same processing modules.
3 Proposed Model for Multilingual Sentiment Analysis
In order for a Tweet to be classified, it has to be represented by a set of words that

expresses its contents. The classification process goes through two different phases:
Training and Testing. Each phase consists of two stages that are different than each
other, as shown in Fig. 1.
Input: Training Corpus Input: Testing Corpus

(Pre-Classified Documents/Tweets) (To be Classified)
Training Testing
Feature Extraction Stage
Pre-Processing Pre-Processing
Feature Vector Feature Vector

Generation Generation
Classification
Classification Stage
Algorithm
Learning Stage
Selection Apply
Classification
Process
Building the
Classification
Model
Output:
Classified Documents/Tweets
Fig. 1. Block diagram of the multilingual sentiment analysis model
In the Training phase, the two stages are Feature Extraction and Learning while in
the Testing phase, the stages are Feature Extraction and Classification. Feature
Selection is described in Subsect. 3.1 and Learning and Classification are presented in
Subsect. 3.2. Though the model is designed generically for any language, in our current
work, we are dealing with Arabic and English languages only.
3.1 The Feature Extraction Stage

The first issue that needs to be addressed in text classification is how to represent texts
and extract their corresponding features. Not only that a proper representation will
make manipulation easier and save processing time and memory, it will also retain the
necessary information without any loss. To represent text documents/Tweets, the
Vector Space Model (VSM) is commonly used. In VSM, each document/Tweet is
represented as a vector or Bag of Words (BoW). That is, each document/Tweet is
represented by the set of words it contains and their frequency regardless of their order
[11, 23, 24].
In our approach, during the pre-processing step, the text documents/Tweets

encoding is examined. Depending on the detected language, each text document/Tweet
is being pre-processed and then undergo a procedure to build their feature vector, as
discussed in details in Sect. 3.1.1. Afterwards, weighted feature vector for each of the
document/Tweet is generated as explained in details in Sect. 3.1.2.
3.1.1 The Pre-processing Step

The Pre-Processing step aims to extract the BoW that represents the textual
document/Tweet by the following tasks:
a. Natural Language Processing (NLP) Parser: The parser detects sentences and
tokenizes document/Tweet for further textual analysis. It performs part of speech
(PoS) tagging to markup each word in the text to a proper tag such as verb, noun,
adjective, etc. in English language [25], and passive verbs, singular noun, proper
noun, etc. in Arabic language [26, 27]. Only verbal and proper noun tagged words
are considered in both languages to reduce the dimension of the extracted feature
vector size.
b. Stopping Words Removal: In natural language, stop words are words that do not
contribute towards the meaning or the idea of what is being communicated, for
examples, a, an, in, at, the, etc. The same concept exists in other languages. In
Arabic language, examples of stopping word are ‫ﺍﻣﺲ‬, ‫ﻭﻟﻢ‬, ‫ﻟﻬﺎ‬, ‫ﻓﻰ‬, …, etc. This is
done by searching for words in a pre-existing list of stop words [28].
c. Data Cleaning and Normalization: In the Arabic language, there are special
characters and words that contains symbols. Continuing towards sentimental
analysis with these special cases may cause ambiguity in the classification process.
In this context, normalizing refers to the replacement of special Arabic characters,
for example, “‫ ”ﺇ‬is replaced with a simple “‫”ﺍ‬. Finally, Tweets with only standard
Arabic characters and numbers are attained. For English datasets, non-English
words and characters containing symbols are removed to attain English alphabets,
numbers, and words, or any combination of them. Furthermore, all capital letters are
replaced with their lowercase ones.
d. Stemming: Compound words are replaced by their morphological root. For
example, the words “tester”, “testing”, and “tested”, all share the same root-word
“test”. Stemming is done by applying any of the stemming algorithms. It is fun-
damental in text analysis to avoid such pattern and redundancy. Thus, different
equivalent morphological forms are replaced by their corresponding root word.
In this paper, for Arabic data, ISRI Arabic stemmer from the nltk python library
was used to perform the stemming process. SRI Arabic stemmer is based on an
algorithm, Arabic stemming without a root dictionary, developed by the Information
Science Research Institute, University of Nevada, Las Vegas, USA. While for English
data, we used Porter English stemmer from the nltk python library to perform the word
stemming process. At the end of the pre-processing phase, we obtain a set of stemmed
bag of words which represents the original feature vector that will be used for the
generation of the feature vector.
3.1.2 Generating a Weighted Feature Vector

The task of generating a weighted feature vector aims to assign weights to each word in
the feature vector to give an indication of their importance. One of the most
straightforward and useful techniques to weigh words is Term Frequency. The limi-
tation of using this technique is that it does not take the length of the document into
account [23]. However, [29, 30] introduce a simple solution to this problem: Term
Frequency-Inverse Document Frequency (TF-IDF). This weight is a statistical measure
used to evaluate how important a word is to a document in a collection or corpus. The
importance increases proportionally to the number of times a word appears in the
document but is offset by the frequency of the word in the corpus.
TF-IDF is composed by two terms: the first computes the normalized Term Fre-
quency (TF) and the second term is the Inverse Document Frequency (IDF).
TF measures how frequently a term occurs in a document. Since every
document/Tweet has a different length, it is possible that a term would appear many
more times in a long document/Tweet than shorter ones [31], therefore,
Number of times term ðtÞ appears in a Tweet

TFðtÞ ¼
Max (of the occurrence of terms in Tweets)
While computing TF, all terms are considered equally important [31]. IDF mea-
sures how important a term is,

Total Number of Tweets
IDFðtÞ ¼ log
Number of Tweets that contains term (t)
Finally, the TF-IDF measure the multiplication of TF and IDF, as follows [31],
TF-IDFðtÞ = TF(t) IDF(t)
This means that larger weights are assigned to terms that appear relatively infre-
quent throughout the corpus, but very frequently in individual document/Tweets.
3.2 The Learning and Classification Stages

In these two stages, the data is processed and prepared in a suitable format. The data is
split into training and testing sets of 5 folds. Then, the training data are processed by
the built in data mining algorithms in Scikit-learn Machine Learning in Python [32].
These algorithms are used in building the classification model for different classifiers
such as Naive-Bayes, Decision Trees, etc. As for the test data, the classification
algorithm corresponding to the built model is applied, and thus classification results are
obtained.
4 Experimental Results and Discussion
In this section, we discuss our experimental setup and the results for evaluating the
performance of our proposed approach on five different datasets. There are two datasets
in Arabic: the Arabic Sentiment Tweets Dataset (ASTD) [2] and the Restaurant
Reviews Dataset (RES) [33]. For English, there are three datasets: the Stanford Twitter
sentiment dataset [34], the Twitter US Airline Sentiment dataset [35] and the Uber Ride
Reviews dataset [36].
4.1 Dataset Characteristics
• The Restaurant Reviews Dataset (RES): The restaurants domain data is taken from
two sources. The first is “Qaym4” with 8.6K Arabic reviews, and the second is
collected from “TripAdvisor” and has 2.6K reviews. Both datasets cover 4.5K
restaurants and have reviews written by over 3 K users.
• Arabic Sentiment Tweets Dataset (ASTD): It contains over 10K Arabic Sentiment
Tweets classified into four classes: subjective positive, subjective negative, sub-
jective mixed, and objective.
• The Stanford Twitter sentiment corpus (Stanford): It contains 1.6 million Tweets
automatically labelled as positive or negative based on emotions.
• The Twitter US Airline Sentiment Dataset (US Airline): This data originally came
from Crowdflower Data for Everyone library. It was scraped from February of 2015
and it has been classified by the contributors as positive, negative, and neutral
Tweets.
• The Uber Ride Reviews Dataset (Uber): It contains 1345 Uber ride reviews that are
collected in the period between 2014–2017.
4.2 Evaluation Criteria

All documents/Tweets for training and testing pass through the stages as discussed in
Sect. 3. Experimental results reported in this section are based on precision, recall, and
F1-measures. The F1-measure is the harmonic mean of precision and recall defined as
follows [37].
2 Precision Recall
F1-measure (Precision, Recall) =
Precision + Recall
In the above formula, precision, and recall are two standard measures widely used
in text categorization literature to evaluate an algorithm’s effectiveness on a given
category where:
True Positive
Precision =
True Positive + False Positive
True Positive
Recall =
True Positive + False Negative
While the accuracy is simply defined by the following equation,
True Positive + True Negative

Accuracy % =
Total # of Positive þ Total # of Negative
4.3 Experimental Results and Discussion

To test the proposed model, experiments were performed on different datasets, two in
Arabic and three in English. As shown in Figs. 2 and 3, the performance of seven
classification algorithms are analyzed. Documents/Tweets datasets were randomly split
into a set of 70% Tweets for training and a remaining set of 30% for testing. 5-folds
cross-validation was used for the evaluation. Table 1 shows the classification accura-
cies, precision, recall, and F-measure for the seven classifiers on the five different
datasets of our model and existing ones in the literature. As can be seen in Fig. 2, for
RES dataset, Multilayer Perceptron classification algorithm performs the best with an
accuracy of 81%. For the ASTD dataset, the Logistic Regression classification algo-
rithm is the most accurate with an accuracy of 68%.
Fig. 2. Accuracies of various classifiers

Fig. 3. F1-scores of various classifiers
In addition, it should be noted that the accuracy of the Bernoulli Naive Bayes
classifier when applied to the US Airline Tweets dataset (76%) and the Uber Ride
Reviews dataset (84%), performs much better than the other classifiers. Furthermore,
the accuracy of Logistic Regression classifier (79%) when applied to the Stanford
Tweets dataset (79%), is higher than the other classifiers.
Moreover, we noted that from the results obtained when applying the kNN clas-
sifier the following:
• The kNN classifier assigns equal weights to each attribute in the feature vector. This
may cause confusion when there are many irrelevant attributes in the data.
• Second, the traditional kNN classifier is based on measuring the distance between
the documents/Tweets using different distance measures, such as Euclidean and
Minkowski, but ignoring the semantic relations between them. This may lead to
misclassification of some of the tested data.
Furthermore, when comparing our obtained results with other results from [1–3],
we find the superiority of our proposed model in all seven classifiers as shown in
Table 1.
Table 1. Results comparison using several classifiers on various datasets

Classification Performance Datasets
algorithm metrics Arabic dataset English dataset
RES ASTD US Airline Uber Stanford
[2] [3] Our [1] Our
results results
Decision tree Accuracy 74 N/A N/A 59 N/A 65 75 68
Precision 0.73 N/A N/A 0.56 0.63 0.65 0.73 0.69
Recall 0.73 N/A N/A 0.58 0.645 0.65 0.73 0.69
F1-score 0.73 N/A N/A 0.57 0.645 0.65 0.73 0.69
Multinomial Accuracy 79 48.4 N/A 67 N/A 70 80 78
Naive Bayes Precision 0.62 N/A N/A 0.6 0.642 0.73 0.84 0.78
Recall 0.79 N/A N/A 0.67 0.647 0.7 0.8 0.78
F1-score 0.69 0.485 N/A 0.55 0.646 0.63 0.72 0.78
Bernoulli Naive Accuracy 77 25.3 N/A 66 N/A 76 84 77
Bayes Precision 0.72 N/A N/A 0.59 0.642 0.77 0.84 0.77
Recall 0.77 N/A N/A 0.66 0.647 0.76 0.84 0.77
F1-score 0.74 0.17 N/A 0.61 0.646 0.77 0.84 0.77
Logistic Accuracy 79 45.1 N/A 68 N/A 74 80 79
regression Precision 0.72 N/A N/A 0.64 0.81 0.74 0.64 0.79
Recall 0.79 N/A N/A 0.68 0.816 0.74 0.8 0.79
F1-score 0.7 0.449 N/A 0.56 0.819 0.71 0.71 0.79
K-nearest Accuracy 74 40.9 N/A 66 N/A 69 84 N/A
neighbors Precision 0.74 N/A N/A 0.48 0.59 0.68 0.82 N/A
Recall 0.74 N/A N/A 0.66 0.592 0.69 0.84 N/A
F1-score 0.74 0.409 N/A 0.54 0.593 0.68 0.81 N/A
Perceptron Accuracy 79 44 N/A 61 N/A 73 80 73
Precision 0.76 N/A N/A 0.59 N/A 0.72 0.78 0.73
Recall 0.79 N/A N/A 0.61 N/A 0.73 0.8 0.73
F1-score 0.77 0.439 N/A 0.6 N/A 0.72 0.79 0.73
Multilayer Accuracy 81 N/A 58.5 63 N/A 74 83 N/A
perceptron Precision 0.78 N/A N/A 0.59 N/A 0.73 0.8 N/A
Recall 0.82 N/A N/A 0.61 N/A 0.74 0.83 N/A
F1-score 0.78 N/A 0.536 0.6 N/A 0.73 0.8 N/A
5 Conclusions and Future Work
This paper addresses the task of performing automatic sentiment analysis over textual
datasets in both Arabic and English languages. To achieve this goal, a classification
model is introduced where textual documents/Tweets pass through a two-stage process.
In the pre-processing step, NLP techniques are applied to get the BoWs that best
represents the datasets, and a textual feature vector is generated. Additionally, TF-IDF
is used to generate the numerical feature vector for classification tasks. After building
the feature vector, different classifiers are used to test various models and their per-
formance. Performance is reported using different metrics: Accuracy, Precision,
F-measure, and Recall. Experimental results through the entire paper indicate that:
• For the RES dataset, the best result is obtained by the Multilayer Perceptron
algorithm with an accuracy of 81%. Also, in terms of true positive rate (TPR), the
best results are obtained by the same classifier with a recall of 82%.
• For the ASTD dataset, the best result was obtained using the Logistic Regression
algorithm with an accuracy of 68%. In terms of TPR, the best results are obtained
by the same classifier with a recall of 68%.
• For the US Airline Tweets dataset, the best result was obtained by the Bernoulli
Naive Bayes algorithm with an accuracy of 76%. But in terms of TPR, the best
result is obtained by the Logistic Regression classifier with a recall of 74%.
• For the Uber Ride Reviews dataset, the best result was obtained by both the Ber-
noulli Naive Bayes algorithm and the kNN classification algorithm with an accu-
racy of 84%. In terms of TPR, the best result was obtained by the same classifiers
with a recall of 84%.
• For the Stanford Tweets dataset, the best result was obtained by Logistic Regression
with an accuracy of 79%. In terms of TPR, the best result was obtained by the same
classifier with a recall of 79%.
Note that kNN and the Multilayer Perceptron classification algorithms do not have
any scores. This is due to hardware limitations.
For future work, there are many avenues to pursue, including:
• Utilize the WordNet ontology for building the extracted feature vector to improve
the classification accuracy.
• Extend and evaluate the proposed system against other datasets in other languages
other than Arabic or English.
• Apply dimensionality reduction techniques to minimize the size of the used feature
vector without affecting the overall system performance.
• Extend the proposed work to handle Tweets that uses Franco-Arabic texts.
• Improve the existing techniques and methodologies to be used in classifying web
text documents.
• Implement a web text document classification system using parallel platforms, such
as Apache Spark, to handle the large amount of data in less time without affecting
the overall system performance.
• Combine the decisions of two or more classification algorithms to improve the
accuracy of the results.
• Perform clustering tasks over the set of feature vectors to help application to give
accurate results and search quickly.
References
1. Rane, A., Anand, K.: Sentiment classification system of Twitter data for US airline service
analysis. In: 2018 IEEE 42nd Annual Computer Software and Applications Conference
(COMPSAC), pp. 769–773. IEEE, Tokyo (2018)
2. Nabil, M., Alay, M., Atiya, A.: ASTD: Arabic sentiment tweets dataset. In: Proceedings of
the 2015 Conference on Empirical Methods in Natural Language Processing, pp. 2515–
2519. Association for Computational Linguistics, Lisbon (2015)
3. Baly, R., Badaro, G., El-Khoury, G., Moukalled, R., Aoun, R., Hajj, H., El-Hajj, W.,
Habash, N., Shaban, K.: A characterization study of Arabic Twitter data with a
benchmarking for state-of-the-art opinion mining models. In: Proceedings of the Third
Arabic Natural Language Processing Workshop (WANLP), pp. 110–118. Association for
Computational Linguistics, Valencia (2017)
4. Heikal, M., Torki, M., El-Makky, N.: Sentiment analysis of Arabic Tweets using deep
learning. Procedia Comput. Sci. 142, 114–122 (2018)
5. Goel, A., Gautam, J., Kumar, S.: Real time sentiment analysis of Tweets using Naive Bayes.
In: 2016 2nd International Conference on Next Generation Computing Technologies
(NGCT). IEEE (2016)
6. Holzinger, A., Stocker, C., Ofner, B., Prohaska, G., Brabenetz, A., Hofmann-Wellenhof, R.:
Combining HCI. In: Natural Language Processing, and Knowledge Discovery-Potential of
IBM Content Analytics as an Assistive Technology in the Biomedical Field, pp. 13–24.
Springer, Heidelberg (2013)
7. Ristoski, P., Paulheim, H.: Semantic web in data mining and knowledge discovery: a
comprehensive survey. Web Semant. Sci. Serv. Agents World Wide Web 36, 1–22 (2016)
8. Mukherjee, S., Shaw, R., Haldar, N., Changdar, S.: A survey of data mining applications and
techniques. Int. J. Comput. Sci. Inf. Technol. (IJCSIT) 6(5), 4663–4666 (2015)
9. Ravi, K., Ravi, V.: A survey on opinion mining and sentiment analysis: tasks, approaches
and applications. Knowl. Based Syst. 89, 14–46 (2015)
10. Neethu, M.S., Rajasree, R.: Sentiment analysis in Twitter using machine learning
techniques. In: 2013 Fourth International Conference on Computing, Communications and
Networking Technologies (ICCCNT), pp. 1–5. IEEE, Tiruchengode (2013)
11. Jindal, R., Malhotra, R., Jain, A.: Techniques for text classification: literature review and
current trends. Webology 12(2), 1–28 (2015)
12. Yang, P., Chen, Y.: A survey on sentiment analysis by using machine learning methods. In:
2017 IEEE 2nd Information Technology, Networking, Electronic and Automation Control
Conference (ITNEC), pp. 117–121. IEEE, Chengdu (2017)
13. Pang, B., Lee, L., Vaithyanathan, S.: Thumbs up? Sentiment classification using machine
learning techniques. In: Proceedings of the ACL-02 Conference on Empirical Methods in
Natural Language Processing (EMNLP), pp. 79–86. Association for Computational
Linguistics, ACM, Philadelphia (2002)
14. Das, S., Behera, R.K., Rath, S.K.: Real-time sentiment analysis of Twitter streaming data for
stock prediction. Procedia Comput. Sci. 132, 956–964 (2018)
15. Kim, S.-B., Han, K.-S., Rim, H.-C., Myaeng, S.H.: Some effective techniques for Naive
Bayes text classification. IEEE Trans. Knowl. Data Eng. 18(11), 1457–1466 (2006)
16. Niu, Z., Yin, Z., Kong, X.: Sentiment classification for microblog by machine learning. In:
2012 Fourth International Conference on Computational and Information Sciences, pp. 286–
289. IEEE, Chongqing (2012)
17. Barbosa, L., Feng, J.: Robust sentiment detection on Twitter from biased and noisy data. In:
Proceedings of the 23rd International Conference on Computational Linguistics: Posters,
COLING 2010, pp. 36–44. ACM, Beijing (2010)
18. Celikyilmaz, A., Hakkani-Tür, D., Feng, J.: Probabilistic model-based sentiment analysis of
Twitter messages. In: 2010 IEEE Spoken Language Technology Workshop, pp. 79–84.
IEEE, Berkeley (2011)
19. Guellil, I., Adeel, A., Azouaou, F., Hachani, A.E., Hussain, A.: Arabizi sentiment analysis
based on transliteration and automatic corpus annotation. In: Proceedings of the 9th
Workshop on Computational Approaches to Subjectivity, Sentiment and Social Media
Analysis, pp. 335–341. Association for Computational Linguistics, Brussels (2018)
20. Bhavitha, B.K., Rodrigues, A.P., Chiplunkar, N.J.: Comparative study of machine learning
techniques in sentimental analysis. In: 2017 International Conference on Inventive
Communication and Computational Technologies (ICICCT), pp. 216–221. IEEE, Coimbat-
ore (2017)
21. Dashtipour, K., Poria, S., Hussain, A., Cambria, E., Hawalah, A.Y., Gelbukh, A., Zhou, Q.:
Multilingual sentiment analysis: state of the art and independent comparison of techniques.
Cogn. Comput. 8(4), 757–771 (2016)
22. Balahur, A., Turchi, M.: Multilingual sentiment analysis using machine translation? In:
Proceedings of the 3rd Workshop on Computational Approaches to Subjectivity and
Sentiment Analysis, pp. 52–60. Association for Computational Linguistics, Jeju (2012)
23. Sedding, J., Kazakov, D.: WordNet-based text document clustering. In: Proceedings of the
3rd Workshop on Robust Methods in Analysis of Natural Language Data, pp. 104–113.
Association for Computational Linguistics, ACM, Geneva (2004)
24. Peng, X., Choi, B.: Document classifications based on word semantic hierarchies. Artif.
Intell. Appl. 5, 362–367 (2005)
25. Marcus, M.P., Marcinkiewics, M.A., Santorini, B.: Building a large annotated corpus of
English: the Penn Treebank. Comput. Linguist. 19(2), 313–330 (1993). Special issue on
using large corpora: II (Association for Computational Linguistics, MIT Press Cambridge)
26. Maamouri, M., Bies, A., Krouna, S., Gaddeche, F., Bouziri, B.: Arabic tree banking
morphological analysis and pos annotation, Ver. 3.8. Linguistic Data Consortium, Univ.
Pennsylvania, Pennsylvania (2009)
27. Albared, M., Omar, N., Ab Aziz, M.J., Nazri, M.Z.A.: Automatic part of speech tagging for
Arabic: an experiment using Bigram hidden Markov model. In: International Conference on
Rough Sets and Knowledge Technology, pp. 361–370. Springer, Heidelberg (2010)
28. NL company, Ranks: Stopword Lists. ranks.nl. (n.d). https://www.ranks.nl/stopwords.
Accessed 19 Jan 2019
29. Wen, C.Y.J.: Text categorization based on a similarity approach. In: Proceedings of
International Conference on Intelligent System and Knowledge Engineering, pp. 1–5.
Atlantis Press, China (2007)
30. Wu, H.C., Luk, R.W.P., Wong, K.F., Kwok, L.: Interpreting TF-IDF term weights as making
relevance decisions. ACM Trans. Inf. Syst. 26(3), 13 (2008)
31. Schütze, H., Manning, C.D., Raghavan, P.: Introduction to Information Retrieval, vol. 39.
Cambridge University Press, Cambridge (2008)
32. Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M.,
Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J.: Scikit-learn: machine learning in
Python. J. Mach. Learn. Res. 12, 2825–2830 (2011)
33. ElSahar, H., El-Beltagy, S.R.: Building large arabic multi-domain resources for sentiment
analysis. In: Gelbukh, A. (ed.) Computational Linguistics and Intelligent Text Processing,
CICLing 2015. LNCS, pp. 23–34. Springer (2015)
34. Go, A., Bhayani, R., Huang, L.: Twitter sentiment classification using distant supervision,
p. 12. CS224N Project Report, Stanford (2009)
35. Twitter-Airline-Sentiment Dataset: taken from the standard Kaggle, vol. 2. Kaggle (2017)
36. Purvank. Uber-rider-reviews-dataset taken from the standard Kaggle, kaggle (2018)
37. Sokolova, M., Lapalme, G.: A systematic analysis of performance measures for classification
tasks. Inf. Process. Manage. 45(4), 427–437 (2009)
The 12th International Workshop on Bio
and Intelligent Computing
(BICom-2019)
A Deep Q-Network with Experience
Optimization (DQN-EO) for Atari’s
Space Invaders
Yan Chen and Elis Kulla(B)
Okayama University of Science, Okayama 700-0005, Japan

t15j066cy@ous.jp, kulla@ice.ous.ac.jp
Abstract. Since Deep Q-Learning was introduced in 2013, a lot of

academia and industry researchers, tried to use it to solve their prob-
lems, in different fields. However, Deep Q-Learning does not consist of
a unified method for solving certain problems. In contrary, every prob-
lem requires specific settings and parameters. In this paper, we propose
a Deep Q-Network with Experience Optimization, for Atari’s “Space
Invaders” environment. Training and Testing results show that, while
using our proposed network, there is a strong correlation between the
key action (shoot) and scoring.
Keywords: Artificial intelligence · Deep Reinforcement Learning ·

Atari · Space Invaders · Experience Optimization
1 Introduction
Reinforcement learning is a machine learning method, which refers to goal-

oriented algorithms, which learn how to attain a complex objective (goal) or
maximize along a particular dimension over many steps; for example, maximize
the points won in a game over many moves. They can start from a blank slate,
and under the right conditions they achieve superhuman performance. These
algorithms are penalized when they make the wrong decisions and rewarded
when they make the right ones, which is also known as reinforcement [3]. Some
algorithms of reinforcement learning, include but are not limited to, Q-Learning,
Deep Q-Network, SARSA, A3C and so on.
Based on the Deep Q-Network Algorithm, the first deep learning model was
introduced in [3], followed by a more advanced model two years later in [4].
It was used in 6 Atari 2600 games from the Arcade Learning Environment [1].
On the other hand, the non-profit company OpenAI introduced a toolkit for
developing and comparing reinforcement learning algorithms, by applying them
in different tasks, from walking to gravity balancing and playing games. The
toolkit is called Gym [2], and provides several environments, where user-defined
learning algorithms can be applied.

https://doi.org/10.1007/978-3-030-15035-8_33
352 Y. Chen and E. Kulla
In this paper we use the Atari environment, and apply our proposed algorithm
in the popular game of “Space Invaders”. The game consists of a ground unit,
which is controlled by the user, which tries to eliminate the spaceships above it,
by shooting laser beams only upwards. The spaceships move sideways, until they
reach the side of the screen, where they step closer to the user-controlled ground
unit. While moving, they also shoot laser beams downward, so the user should
avoid two situations; being hit by the laser beams and letting the spaceships reach
the ground. The user scores points for each eliminated spaceship, depending on
the position of the spaceship (the highest the ship, the highest the score). If the
ground unit is eliminated three times, the game is over, and the accumulated
score is the final result of the game.
In this work, we propose a Deep Q-Network with Experience Optimization
(DQN-EO) algorithm. We use the Space Invaders’ (OpenAI’s Gym) environment
and apply our algorithm in order to evaluate its performance, regarding reward
drop rate, total episode score and episode lifetime. The remainder of the paper
is organized as follows.
In Sect. 2, we summarize the basics of Reinforcement learning, and how it
is applied to the Space Invaders environment, which is presented in Sect. 3. In
Sect. 4, we describe the four phases of our proposed DQN-EO algorithm; experi-
ence optimization, search algorithms and neural network structure. Training of
the algorithm is explained in Sect. 5. Then, we discuss the results in Sect. 6 and
conclude our paper in Sect. 7.
2 Reinforcement Learning
In Reinforcement Learning, the agent gets a state s from environment env and
selects an action a to apply to the environment env. The environment will feed-
back a reward (Qπ(s, a)) to the agent by evaluating the action and its current
state (see Fig. 1). In the future the agent will be oriented to choose the action
which previously obtained the highest reward. Different reinforcement learning
algorithms have been developed and applied to different problems. To name a
few, Q-Learning, DQN, SARSA, DDPG are widely-used ones. In Q-Learning,
agents use Q-tables to store the Q-value for every action-state pair, that it ever
experienced. If agent gets a reward r in time t, Q-Learning will update Q-table
so that st will get the full reward r, st−1 , st−2 . . . st−n will get smaller rewards as
defined by the discount factor γ, respectively, r × γ, r × γ 2 . . . r × γ n . Q-Learning
solves simple problems quickly, but it also falls into an infinite loop easily. In
complicated problems, the number of possible states is so big, that it is not very
practical to store all of them in a Q table, because searching for an existing state
in a huge Q-table will consume valuable resources. Deep Q-Network (DQN) uses
Neural Networks together with Q-Learning, to solve this problem. DQN uses
states as Neural Networks’ input, with several hidden layers. Neural Networks
will calculate the reward Q for all possible actions and choose to apply the action
with the highest reward to the environment. However, feedback about an action
can only be received after several time-steps of the running algorithm, and the
A Deep Q-Network with Experience Optimization (DQN-EO) 353
effect of key actions (fire, move, jump) to the reward (score) is delayed and not
accurate. We will try to solve this problem in this paper.
Fig. 1. Agent and environment
In Table 1, we shortly summarize the terminology used in reinforcement learn-

ing algorithms.
In this paper, we use the networks in NIPS 2013 [3] and Nature 15 [4], and
change the search methods to our proposed algorithm, in order to compare its
performance.
3 Environment
We use TensorFlow as Deep Reinforcement Learning API and OpenAI GYM

as environment and emulator in our experiments. OpenAI GYM, among others
contains an emulator for many games from Atari 2600, which was a very popular
game console in 1970s. It used an 8-bit MOS Technology 6507 CPU @ 1.19 MHz
with 128 bytes RAM. The emulator is provided by an Arcade Learning Environ-
ment [ALE]. We use “SpaceInvaders-v0” as the environment in our experiments.
The environment accepts 6 values of actions (0 5) as input (see Table 3). As
environment state, we get 4 values, as shown in Table 2. (1) a 210 × 160 pixel
images with a 128 color palette, (2) an integer score, (3) a gameover flag and (4)
the number of lives remaining.
4 Proposed DQN-EO Algorithm
Our proposed DQN with Experience Optimization Algorithm consists of 3

important parts: Experience Optimization, Search Algorithm and Neural Net-
work. We will explain them in the following.
Table 1. Reinforcement learning terminology
Terminology Description
Agent – The entity that takes action a (see below)
on the environment env (see below)
Action (a) – Chosen from a list of possible actions,
depending on application
Environment (env) – The “world” in which the agent moves,
takes action and gets rewarded or penalized
State (s) – A location, situation or condition of the
agent in a certain time
Reward (r) – The feedback of the environment for
successful agent’s actions. A penalty when
the action fails
Discount factor (γ) – A factor multiplied to future rewards,
which makes future rewards worth less than
immediate rewards
Policy (π) – A strategy, which maps states to actions,
in order for the agent to determine its next
action based on the current state
Value (V π(s)) – The expected future reward multiplied by
the discount factor (see above). It is defined
as the expected reward of the current state
under policy (π)
Q-value or action value – Similar to value v, but it takes action a as
(Qπ(s, a)) a parameter
Table 2. Environment state data
Input Description Value

Image 210px × 160px 24bits/px
Score Integer 0–200
Lives Remaining lives 0–3
GOF GameOver flag 0 or 1
Table 3. Action mapping
# Original Remapped Reduce 3 (%3)

0 Do nothing Do nothing 0
1 Fire Right 1
2 Right Left 2
3 Left Fire 0
4 Right + Fire Right + Fire 1
5 Left + Fire Left + Fire 2
4.1 Experience Optimization

In order to do that, the algorithm follows the following steps.
1. Remap actions as in Table 3. For repetitive actions we will use its reduced
value, in order to minimize the number of fire action data.
2. Get an initial state from the environment.
3. 47 actions will be stored in a buffer.
4. Randomly generate an action from -greedy policy.
5. Get reward (score) from environment.
6. Search (as explained in Sect. 4.2) for fire action in the buffer, and update its
reward.
7. Push last 47 data from buffer to experience base D.
Table 4. Action examples
Original Action 3 3 3 3 5 5 5 5 4 4 4 4 2 2 2 2
Reduced Action 3 0 0 0 5 2 2 2 4 1 1 1 2 2 2 2
Search Space b × a × c
4.2 Search Algorithm

In Space Invaders, and other similar situations, it is very difficult to relate a
certain action with the score obtained from eliminating the ships, because the
ships’ movement speed increases and the distance from the ground changes con-
tinuously. Thus, we need to search backwards for an undefined number of frames.
We collected 1000 records for every positive score (5, 10, 15, 20, 25, 30, 200)
for a total of 7000 records. From 7000 records, we find out relation between the
actual frame number x in the episode and the approximation of the number
of frames y that passed since the fire action until each score. For example, if
agent got a score of 5, and we are playing the game on its 1000’s frame, it has
a 99% probability that fire frame was in a range of int(−0.0043 × x + 8.9661) ±
2 = int(4.6661 ± 2) = 4 ± 2 (see Fig. 2 for other scores). Because this is an
approximation, in our buffer we search for the fire frame in the following order
4, 3, 5, 2, 6. If fire frame doesn’t exist in this range, the current reward will be
dropped.
Because of action mapping and reducing, every 4 frames would have only
1 fire at most. In Table 4, lets consider that the agent fired in frame a. If the
prediction from our method above falls in the green (filled) squares around a,
then the search would get the correct answer, a. Only if the prediction falls in
the red squares (marked with ×), the search will mistakenly identify action b or
action c as the responsible fire actions. But, this is the worst scenario, because
it is very rare that the agent fires 3 times consecutively (if is 1, probability is
12.5%). Furthermore, after adequate training, the agent is expected to fire in a
better timing so consecutive fire actions might disappear.
4.3 Neural Networks
We build 2 types of networks for training our data. One network is similar to that
of NIPS 2013 (Net13). The input to the neural network consists of a 84 × 84 × 4
image produced from the experience base D. In the first hidden layer 16 filters
are used. Each of them has 8 × 8 size with stride 4 with the input image and
applies a nonlinear rectifier. The second hidden layer uses 32 filters. Each of them
has a 4 × 4 size with stride 2, again followed by a nonlinear rectifier. The final
hidden layer is fully-connected and consists of 256 rectifier units. The output
layer is a fully-connected linear layer with a single output for each valid action
(6 outputs).
The other one is the same as the network in Nature 2015 (Net15). The
input to the neural network consists of a 84 × 84 × 4 image produced from the
experience base D. In the first hidden layer 16 filters are used. Each of them has
8 × 8 size with stride 4 with the input image and applies a nonlinear rectifier. The
second hidden layer uses 32 filters. Each of them has a 4 × 4 size with stride 2,
again followed by a nonlinear rectifier. This is followed by a third convolutional
layer that uses 64 filters of 3 × 3 with stride 1 followed by a rectifier. The final
hidden layer is fully-connected and consists of 512 rectifier units. The output
layer is a fully-connected linear layer with a single output for each valid action
(6 outputs).
Fig. 2. Relation between the actual frame number x in the episode and the approx-
imation of the number of frames y that passed since the fire action until each score
5 Training
We will train 2 algorithms (the original from NIPS 2013 and the proposed DQN-
OE) with each Network. As rectifier, which is also known as activation function,
we use ReLU (Rectified Linear Unit), as shown in Figs. 3 and 4.
For simplicity, and without deteriorating the performance, during training,
we changed all images to 84 × 84 grayscale, all positive rewards to 1, And all
negative rewards to −1. In this case, Algorithm 2 will receive negative reward
every time the number of lives decreases, 3 times more than Algorithm 1, which
will only receive negative reward when the game is over (number of lives =
0). So, We added the negative reward in Algorithm 1 for fair comparison. We
used AdamOptimizer provided by TensorFlow with minibatches of size 32 and
an -greedy policy. value was initialized to 0.9, and programmed to reduce to
0.1 linearly over 1 million training steps. It will be fixed at 0.1 after 1 million
steps. We trained for a total of 60 million steps and used a replay memory
of 100 thousand steps. The algorithm selects an action every 4 frames, and
emulator will repeat this action in the next 3 frames. Training program will be
active every 4 frames, similar to action selector. We use the actual frame and
3 previous frames as input for the networks. Thus, 1 step actually considers 4
frames in total. The trainer will output model files every 10 thousand steps for
a total of 599 models after training. When trainer successfully outputs a model
Fig. 3. NetWorks in NIPS 2013
Fig. 4. NetWorks in nature 2015

or the reward is dropped, the algorithm will record the reward drop rate and
the reward, which are used for evaluation purposes.
6 Result
After completing the training process, we made 4 evaluators (2 networks with

each algorithm) to run every model in 100 episodes and log out drop rate, aver-
age score, maximum score and frames spent during 100 episodes (lifetime). After
completing the evaluation a model, the evaluators will load next model file auto-
matically and the value of will be fixed at 0.05 during the evaluation.
6.1 Drop Rate
The Drop Rate, shows how often a reward is not counted towards training the
systems, because the search algorithm could not find a fire action in the buffer.
In Fig. 5, we can see that at the beginning there are 2.5% drops, but after 2.5
million train steps, the drop rate reduces to under 1%.
Fig. 5. Drop rate
6.2 Score
The score shows, the total score that the agent got from each episode. In Fig. 6,
the average and maximum score for 4 different cases. We can notice that, in the
beginning of the training, DQN-EO got much higher average score and higher
maximum score than the original one. However when decreased to 0.1 at 1
million steps, DQN-EO’s performance decreased. In Net13, DQN-EO caught up
original agent at step 3 million and then had the same performance with the
original one. In Net15, DQN-EO was better after 2.5 million steps. Moreover it
got the highest average score (457) and highest score (1275) from all the cases.
Fig. 6. Average score and maximum score
Fig. 7. Frames spent during 100 episodes
6.3 Lifetime
In Fig. 7, lifetime of DQN-EO was higher than the original one. Higher lifetime
shows that agent with Experience Optimization was better in avoiding enemies’
laser.
7 Conclusions
In this paper, we proposed a Deep Q-Network with Experience Optimization, for

Atari’s “Space Invaders” environment. Results from training and testing show
that, there was a strong correlation between shooting and scoring. Among all
tested cases, DQN-EO improved the average and maximum score, which shows
the timing of the shoot was better. Moreover, the agent learned how to better
avoid the enemy lasers, which is shown by the improved lifetime results.
References
1. Bellemare, M.G., Naddaf, Y., Veness, J., Bowling, M.: The arcade learning environ-
ment: an evaluation platform for general agents. J. Artif. Intell. Res. 47, 253–279
(2013)
2. OpenAI Gym Toolkit. https://gym.openai.com/
3. Mnih, V., Kavukcuoglu, K., Silver, D., Graves, A., Antonoglou, I., Wierstra, D.,
Riedmiller, M.: Playing Atari with Deep Reinforcement Learning. arXiv, 19 Decem-
ber 2013. https://arxiv.org/pdf/1312.5602v1.pdf
4. Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A., Veness, J., Bellemare, M.G.,
Graves, A., Riedmiller, M., Fidjeland, A.K., Ostrovski, G., Petersen, S., Beattie,
C., Sadik, A., Antonoglou, I., King, H., Kumaran, D., Wierstra, D., Legg, S., Has-
sabis, D.: Human-level control through deep reinforcement learning. Nature. 518,
529–33 (2015). https://doi.org/10.1038/nature14236
Design of a Deep Q-Network Based
Simulation System for Actuation Decision
in Ambient Intelligence
Tetsuya Oda(B) , Chiaki Ueda, Ryo Ozaki, and Kengo Katayama
Department of Information and Computer Engineering, Okayama University

of Science (OUS), 1-1 Ridaicho, Kita-ku, Okayama 700–0005, Japan
oda.tetsuya.fit@gmail.com, {ueda,ozaki,katayama}@ice.ous.ac.jp
Abstract. Ambient Intelligence (AmI) deals with a new world of ubiq-

uitous computing devices, where physical environments interact intelli-
gently and unobtrusively with people. AmI environments can be diverse,
such as homes, offices, meeting rooms, schools, hospitals, control centers,
vehicles, tourist attractions, stores, sports facilities, and music devices.
This paper presents design and implementation of a simulation system
based on Deep Q-Network (DQN) for actuation decision in AmI. DQN
is a deep neural network structure used for estimation of Q-value of the
Q-learning method. We implemented the proposed simulating system by
Rust programming language. We describe the design and implementation
of the simulation system, and show some simulation results to evaluate
its performance.
1 Introduction
Ambient Intelligence (AmI) is the vision that technology will become invisible,
embedded in our natural surroundings, present whenever we need it, enabled
by simple and effortless interactions, attuned to all our senses, adaptive to users
and context and autonomously acting [1,2]. High quality information and content
must be available to any user, anywhere, at any time, and on any device.
In order that Ambient Intelligence becomes a reality, it should completely
envelope humans, without constraining them. Distributed embedded systems for
Ambient Intelligence are going to change the way we design embedded systems,
in general, as well as the way we think about such systems. But, more impor-
tantly, they will have a great impact on the way we live. Applications ranging
from safe driving systems, smart buildings and home security, smart fabrics or e-
textiles, to manufacturing systems and rescue and recovery operations in hostile
environments, are poised to become part of society and human lives.
There are a lot of works done on testbed for AmI. In [3], the authors present
a simulation environment that offers a library of Networked Control Systems
(NCS) blocks. Thus, the constraints can be considered and integrated in the
design process. They describe a real process, an inverted pendulum, which is

https://doi.org/10.1007/978-3-030-15035-8_34
Performance Evaluation of an AmI Testbed for Improving QoL 363
automated based on Mica nodes. These nodes were designed especially for Ambi-
ent Intelligence purposes. This real NCS serves as a challenging benchmark for
proving the AmI suitability of the controllers.
In [4], the authors present the development of an adaptive embedded agent,
based on a hybrid PCA-NFS scheme, able to perform true real-time control
of Ambient Intelligence environments in the long term. The proposed architec-
ture is a single-chip HW/SW architecture. It consists of a soft processor core
(SW partition), a set of NFS cores (HW partition), the HW/SW interface, and
input/output (I/O) peripherals. An application example based on data obtained
in an ubiquitous computing environment has been successfully implemented
using an FPGA of Xilinx’s Virtex 5 family [5].
In [6], the authors describe a framework to Context Acquisition Services and
Reasoning Algorithms (CASanDRA) to be directly consumed by any type of
application needing to handle context information. CASanDRA decouples the
acquisition and inference tasks from the application development by offering a
set of interfaces for information retrieval. The framework design is based on a
data fusion-oriented architecture. CASanDRA has been designed to be easily
scalable; it simplifies the integration of both new sensor access interfaces and
fusion algorithms deployment, as it also aims at serving as a testbed for research.
There are many critical issues arising in AmI. Based on the specific applica-
tion, different objectives can be taken into account such as energy consumption,
throughput, delay, coverage, etc. [7,8]. Also many schemes have been proposed
in order to optimize a specific Quality of Service (QoS) parameter. We propose
a simulation system based on Deep Q-Network (DQN) as a controller for actor
node mobility and implement by Rust programming language [9,10]. We describe
the implementation and show the simulation results of the simulation system.
The rest of the paper is organized as follows. In Sect. 2, we describe the basics
of Ambient Intelligence. In Sect. 3, we present the overview of DQN. In Sect. 4,
we show the description and design of the simulation system. Simulation results
are shown in Sect. 5. Finally, conclusions and future work are given in Sect. 6.
2 Ambient Intelligence
In the future, small devices will monitor the health status in a continuous man-
ner, diagnose any possible health conditions, have conversation with people to
persuade them to change the lifestyle for maintaining better health, and com-
municates with the doctor, if needed [11]. The device might even be embedded
into the regular clothing fibers in the form of very tiny sensors and it might
communicate with other devices including the variety of sensors embedded into
the home to monitor the lifestyle. For example, people might be alarmed about
the lack of a healthy diet based on the items present in the fridge and based on
what they are eating outside regularly.
The Ambient Intelligence paradigm represents the future vision of intelligent
computing where environments support the people inhabiting them [12–14]. In
this new computing paradigm, the conventional input and output media no
364 T. Oda et al.
Algorithm 1. Deep Q-learning with Experience Replay

1: Initialize replay memory D to capacity N
2: Initialize action-value function Q with random weights
3: for episode = 1, M do
4: Initialise sequence s1 = {x1 } and preprocessed sequenced φ1 = φ(s1 )
5: for t = 1, T do
6: With probability ε select a random action at
7: otherwise select at = maxa Q∗ (φ(st ), a; θ)
8: Execute action at in emulator and observe reward rt and image xt+1
9: Set st+1 = st , at , xt+1 and preprocess φt+1 = φ(st+1 )
10: Store transition (φt , at , rt , φt+1 ) in D
11: Sample random
minibatch of transitions (φj , aj , rj , φj+1 ) from D
rj for terminal φj+1
12: Set yj =
rj + maxa (φj+1 , aj ; θ) for non-terminal φj+1
13: Perform a gradient descent step on (yj - Q(φj , aj ; θ))2 according to equation 3
14: end for
15: end for
longer exist, rather the sensors and processors will be integrated into everyday
objects, working together in harmony in order to support the inhabitants [15]. In
these environments, a multitude of interconnected, invisible embedded systems,
seamlessly integrated into the background, surround the user [16]. The system
programs to meet their needs by recognizing people living in it and learning
from action. By relying on various artificial intelligence techniques, Ambient
Intelligence promises the successful interpretation of the wealth of contextual
information obtained from such embedded sensors, and will adapt the environ-
ment to the user needs in a transparent and anticipatory manner.
As necessary for Ambient Intelligence, media management and processing,
natural interaction, computer intelligence, emotion calculation etc. are included.
Technology that can accurately judge surrounding conditions by machine learn-
ing or artificial intelligence is necessary. Thereby, it is possible to flexibly pro-
vide the most suitable service at that time. Artificial intelligence is important to
achieve AmI and it is thought that it is necessary to receive more information
from artificial intelligence [17].
3 DQN
The algorithm for training Deep Q-learning is presented in Algorithm 1. The

DQN defined the tasks between the agents and the environments [18–21] in
Atari 2600 games emulator [22]. Because Q maps history-action pairs to scalar
estimates of their Q-value, the history and the action have been used as inputs
to the neural network by some previous approaches [23,24]. The environment
was set as ξ. At each step, the agent selected an action at from the action sets
of the game and observed a displayed image xt from the current screen [25].
The change of the game score rt was regarded as the reward for the action. For
a standard reinforcement learning method, we can complete all of these game
sequences st as Markov decision process directly, where sequences of actions
and observations st = x1 , a1 , x2 , . . . , at−1 , xt . Also, it uses a technique known as
experience replay [26] in which it store the agent’s experiences at each time-step,
et = (st , at , rt , st+1 ) in a data-set D = e1 , . . . , eN , pooled over many episodes
into a replay memory. Defining the discounted reward for the future Tby atfactor
−t
γ, the sum of the future reward until the end would be Rt = t =t γ rt .
T means the termination time-step of the game. After performing experience
replay, the agent selects and executes an action according to an ε-greedy policy.
Since using histories of arbitrary length as inputs to a neural network can be
difficult, Q-function of DQN instead works on fixed length representation of
histories produced by a function φ. The target was to maximize the action-value
function Q∗ (s, a) = maxπ E[Rt |st = s, at = a, π], where π is the strategy for
choosing of best action. From the Bellman equation, it is equal to maximize the
expected value of r + γQ∗ (s , a ), if the optimal value Q∗ (s , a ) of the sequence
at the next time step is known.
Q∗ (s , a ) = Es ∼ξ [r + γmax

Q∗ (s , a )|s, a] (1)
a
Not using iterative updating method to optimal the equation, it is common

to estimate the equation by using a function approximator. Q-network in DQN
was such a neural network function approximator with weights θ and Q(s, a; θ) ≈
Q∗ (s, a). The loss function to train the Q-network is:
Li (θi ) = Es,a∼ρ(· ) [(yi − Q(s, a; θi ))2 ]. (2)
The yi is the target, which is calculated by the previous iteration result θi−1 .
ρ(s, a) is the probability distribution of sequences s and a. The gradient of the
loss function is shown in Eq. (3):
∇θi Li (θi ) = Es,a∼ρ(· );s ∼ξ [(yi − Q(s, a; θi ))∇θi Q(s, a; θi )]. (3)
4 Design and Implementation of Proposed Simulation

System
In this section, we present design and implementation of proposed simulation
system based on DQN for actuation decision in AmI. The simulation system
structure is shown in Fig. 1. The proposed simulating system is implemented by
Rust programming language [27,28]. Rust is a system programming language
focused on three goals: safety, speed, and concurrency [29]. Rust supports a
mixture of programming styles: imperative procedural, concurrent actor, object-
oriented and functional.
We consider tasks in which an agent interacts with an environment. In this
case, the actuation node of illumination give visual environment step by step
in a sequence of actions, observations and rewards. We took in consideration
the uniformity and lights]out time of illumination in ergonomics of actuator
nodes [30].
366 T. Oda et al.
Fig. 1. The structure of DQN based actuation decision simulation system.
The actuator nodes consider RGB LED. In order to decide the reward func-
tion, we considered the Uniformity (U) and Working Hours (WH) parameters.
The reward function r is defined as follows:

5 × (U + W H) (if staten ≥ statemax )
r= (4)
−1 × (U + W H) (else) .
Table 1. Simulation parameters.
Parameters Values
Number of episode 100000
Number of iteration 100
Number of hidden layers 3
Number of hidden units 15
Initial weight value Normal initialization
Activation function ReLU
Action selection probability (ε) 0.999 − (t/Number of episode)
(t = 0, 1, 2, . . ., Number of episode)
Learning rate (α) 0.04
Discount rate (γ) 0.9
Experience memory size 300 × 100
Batch size 32
The initial weights values are considered as Normal Initialization [31]. The
input layer is using the position of events and actor nodes, mobile actor node
patterns and total reward values in Experience Memory. The hidden layer is
connected with 256 rectifier units in Rectified Linear Units (ReLU) [32]. The
output Q-values are mobile actuator node patterns.
Table 2. Types of sensing and actuation.
Types of sensing
Type Values
Body motion left, right, up, down, stationary
Emotion delight, anger, sorrow, pleasure
Types of actuation
Type Values
Uniformity up, down, stationary
Light color white, red, blue, green, lights]out
The simulation parameters are shown in Tables 1 and 2. One episode has 200
iterations.
368 T. Oda et al.
Table 3. Results of total reward.
Episode Total reward

Best episode 7590
Median episode 5265
Worst episode −1525
30000
20000
10000
Reward
-10000
-20000
Best
Median
Worst
-30000
0 50 100 150 200
Number of Iteration
Fig. 2. Results of reward.
In Table 3, we show the simulation total reward of episodes. There are best,
median and worst episodes. The best episode shows that actuator nodes are good
affect for visual environment. In Fig. 2, we show the simulation results of reward
vs. number of iteration. The higher reward value means that the actuator node
can effect visual environment.
6 Conclusion
In this work, we designed a simulation system based on DQN for actuation

decision in AmI. We presented the implementation of the proposed simulation
system and have shown also the interface in a simulation scenario.
In the future, we would like to make extensive simulations for different sim-
ulation scenarios.
Acknowledgement. This work was supported by Faculty of Engineering, Okayama

University of Science (OUS) Grant-in-Aid for Exploratory Research.
References
1. Lindwer, M., Marculescu, D., Basten, T., Zimmermann, R., Marculescu, R., Jung,
S., Cantatore, E.: Ambient intelligence visions and achievements: linking abstract
ideas to real-world concepts. In: Design, Automation and Test in Europe Confer-
ence and Exhibition, pp. 10–15 (2003)
2. Obukata, R., Cuka, M., Elmazi, D., Oda, T., Ikeda, M., Barolli, L.: Design and
evaluation of an ambient intelligence testbed for improving quality of life. Int. J.
Space-Based Situated Comput. 7(1), 8–15 (2017)
3. Gabel, O., Litz, L., Reif, M.: NCS testbed for ambient intelligence. In: IEEE Inter-
national Conference on Systems, Man and Cybernetics, vol. 1, pp. 115–120 (2005)
4. del Campo, I., Martinez, M.V., Echanobe, J., Basterretxea, K.: A hard-
ware/software embedded agent for realtime control of ambient-intelligence envi-
ronments. In: IEEE International Conference on Fuzzy Systems (FUZZ-IEEE),
pp. 1–8 (2012)
5. Virtex 5 Family Overview, Xilinx Inc., San Jose, CA (2009)
6. Bernardos, A.M., Tarrio, P., Casar, J.R.: CASanDRA: a framework to provide con-
text acquisition services and reasoning algorithms for ambient intelligence applica-
tions. In: International Conference on Parallel and Distributed Computing, Appli-
cations and Technologies, pp. 372–377 (2009)
7. Costanzo, C., Loscri, V., Natalizio, E., Razafindralambo, T.: Nodes self-deployment
for coverage maximization in mobile robot networks using an evolving neural net-
work. Comput. Commun. 35(9), 1047–1055 (2012)
8. Li, Y., Li, H., Wang, Y.: Neural-based control of a mobile robot: a test model for
merging biological intelligence into mechanical system. In: IEEE 7th Joint Interna-
tional Information Technology and Artificial Intelligence Conference (ITAIC-2014),
pp. 186–190 (2014)
9. Oda, T., Obukata, R., Ikeda, M., Barolli, L., Takizawa, M.: Design and imple-
mentation of a simulation system based on deep Q-network for mobile actor node
control in wireless sensor and actor networks. In: The 31th IEEE International Con-
ference on Advanced Information Networking and Applications Workshops (IEEE
WAINA-2017) (2017)
10. Oda, T., Elmazi, D., Cuka, M., Kulla, E., Ikeda, M., Barolli, L.: Performance evalu-
ation of a deep Q-network based simulation system for actor node mobility control
in wireless sensor and actor networks considering three-dimensional environment.
In: The 9th International Conference on Intelligent Networking and Collaborative
Systems (INCoS-2017), pp. 41–52 (2017)
11. Acampora, G., Cook, D., Rashidi, P., Vasilakos, A.V.: A survey on ambient intel-
ligence in health care. Proc. IEEE 101(12), 2470–2494 (2013)
12. Aarts, E., Wichert, R.: Ambient Intelligence. Technology Guide, pp. 244–249 (2009)
13. Aarts, E., de Ruyter, B.: New research perspectives on ambient intelligence. J.
Ambient Intell. Smart Environ. 1(1), 5–14 (2009)
14. Vasilakos, A., Pedrycz, W.: Ambient Intelligence, Wireless Networking, and Ubiq-
uitous Computing. Artech House, Norwood (2006)
15. Sadri, F.: Ambient intelligence: a survey. ACM Comput. Surv. 43(4), 36:1–36:66
(2011)
370 T. Oda et al.
16. Hagras, H., Callaghan, V., Colley, M., Clarke, G., Pounds-Cornish, A., Duman,
H.: Creating an ambient-intelligence environment using embedded agents. IEEE
Intell. Syst. 19(6), 12–20 (2004)
17. Ramos, C., Augusto, J.C., Shapiro, D.: Ambient intelligence-the next step for arti-
ficial intelligence. IEEE Intell. Syst. 23(2), 15–18 (2008)
18. Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A.A., Veness, J., Bellemare, M.G.,
Graves, A., Riedmiller, M., Fidjeland, A.K., Ostrovski, G., Petersen, S., Beattie,
C., Sadik, A., Antonoglou, I., King, H., Kumaran, D., Wierstra, D., Legg, S.,
Hassabis, D.: Human-level control through deep reinforcement learning. Nature
518, 529–533 (2015)
19. Mnih, V., Kavukcuoglu, K., Silver, D., Graves, A., Antonoglou, I., Wierstra, D.,
Riedmiller, M.: Playing atari with deep reinforcement learning. arXiv:1312.5602v1,
pp. 1–9 (2013)
20. Oda, T., Kulla, E., Cuka, M., Elmazi, D., Ikeda, M., Barolli, L.: Performance
evaluation of a deep Q-network based simulation system for actor node mobility
control in wireless sensor and actor networks considering different distributions of
events. In: The 11th International Conference on Innovative Mobile and Internet
Services in Ubiquitous Computing (IMIS-2017), pp. 36–49 (2017)
21. Oda, T., Kulla, E., Katayama, K., Ikeda, M., Barolli, L.: A deep Q-network based
simulation system for actor node mobility control in WSANs considering three-
dimensional environment: a comparison study for normal and uniform distribu-
tions. In: The 12th International Conference on Complex, Intelligent, and Software
Intensive Systems, pp. 842–852 (2018)
22. Lei, T., Ming, L.: A robot exploration strategy based on Q-learning network. In:
IEEE International Conference on Real-Time Computing and Robotics (RCAR-
2016), pp. 57–62 (2016)
23. Riedmiller, M.: Neural fitted Q iteration - first experiences with a data efficient neu-
ral reinforcement learning method. In: The 16th European Conference on Machine
Learning (ECML-2005), vol. 3720 of the series Lecture Notes in Computer Science,
pp. 317–328 (2005)
24. Lange, S., Riedmiller, M.: Deep auto-encoder neural networks in reinforcement
learning. In: The 2010 International Joint Conference on Neural Networks (IJCNN-
2010), pp. 1–8 (2010)
25. Kaelbling, L.P., Littman, M.L., Cassandra, A.R.: Planning and acting in partially
observable stochastic domains. Artif. Intell. 101(1–2), 99–134 (1998)
26. Lin, L.J.: Reinforcement learning for robots using neural networks. Technical
report, DTIC Document (1993)
27. The Rust Programming Language (2007). https://www.rust-lang.org/
28. GitHub - rust-lang/rust: a safe, concurrent, practical language (2007). https://
github.com/rust-lang/
29. ‘rust’ tag wiki - Stack Overflow (2007). http://stackoverflow.com/tags/rust/info/
30. Ohno, Y.: Color rendering and luminous efficacy of white LED spectra. In: The
4th International Conference on Solid State Lighting, pp. 88–98 (2004)
31. Glorot, X., Bengio, Y.: Understanding the difficulty of training deep feedforward
neural networks. In: The 13th International Conference on Artificial Intelligence
and Statistics (AISTATS-2010), pp. 249–256 (2010)
32. Glorot, X., Bordes, A., Bengio, Y.: Deep sparse rectifier neural networks. In: The
14th International Conference on Artificial Intelligence and Statistics (AISTATS-
2011), pp. 315–323 (2011)
An Efficient Scheduling of User
Appliances Using Multi Objective
Optimization in Smart Grid
Hafiz Muhammad Faisal1 , Nadeem Javaid1(B) , Umar Qasim2 , Shujaat Habib3 ,

Zeshan Iqbal4 , and Hasnain Mubarak4
1
Comsats University Islamabad, Islamabad 44000, Pakistan
2
Cameron Library, University of Alberta, Edmonton, AB T6G 2J8, Canada
3
Air University, Multan, Pakistan
4
Abstract. Electricity is the basic demand of consumers. With the pas-

sage of time this demand is increasing day by day. Smart grid (SG) trying
to fulfill the demand of customers. When demand increases then load is also
high. To maintain load from on peak hours to off peak hours, consumer
needs to manage their appliances by home energy management system
(HEMS). HEMS schedule the appliances according to customer’s needs. In
this paper, scheme is proposed which is used to minimize the electricity cost
and also maximize the user comfort. The proposed scheme is performed
better than existing meta heuristic techniques. The proposed scheme is
used real time price (RTP) price signal. Simulation results shows that the
algorithm has met the objective of DSM. Moreover, the proposed algo-
rithm outperforms earth worm algorithm (EWA) and single swam opti-
mization (SSO) in terms of electricity cost and user comfort.
Keywords: Smart grid · Home energy management system ·

Real time price
1 Introduction
Energy is one of the most important resource, and energy demand is grow-
ing every day. Service providers are facing many problems to fulfill the energy
demand in residential building and industrial sectors. There are two ways to
solve this problem.
1. Produce additional energy and find new resources to produce energy
2. Excellent usage of existing resources
The first approach is costly and time consuming, as compared to the sec-
ond approach that is more efficient and inexpensive. Information technology
evolved and many schemes are introduced for energy consumption optimization.
https://doi.org/10.1007/978-3-030-15035-8_35
372 H. M. Faisal et al.
The bi-directional communication is not only concerned by the consumers for

electricity price and maintenance schedules of the distribution network, however,
also motivate the providers to monitor and analyze the real time power utiliza-
tion data. Smart meters (SM) are installed in the residential area. SM provides
the user’s complete information of demand, supply and price signal. The energy
consumption is increasing at a rapid rate in residential areas, hence the effi-
cient use of energy is a big issue in the residential sectors. DSM has two main
functions; load management and demand response (DR). DR is one of the core
function of DSM. DR can be termed as the series of steps taken by customer
in reaction to the changing price rates announced by utility. Due to rapidly
changing grid conditions demand level can also be changed. This varying change
causes a mismatch between demand and supply. This mismatch is dangerous for
the integrity of grid that is spread over a large area. That is the reason DR is
used as it provides flexibility at relatively low rates. DR always try to adjust
the power demand of consumers. The DR scheme helps the customers to save
electricity bills when the prices are high in peak hours. Customers can shift the
usage of their own appliances into off peak hours. Many researchers proposed
different schemes in the literature. DR is divided into incentive based programs
and price based programs. The utility can control the appliances of the user
and provides the financial incentives for demand reduction. However, privacy is
compromised by directly accessing the appliances of customers. In price based
programs, end users change the power consumption in their houses according to
the price schemes which is provided by the utility. The power generation and
power utilization is a one way process so, the power generation system is unable
to control and manage electricity consumption. SM provides two way communi-
cation between user and utility. Smart home (SH) and SM are very important
in residential building for reducing energy consumption due to information com-
munication technology advancement. In this paper, DSM technique implements
for scheduling the appliances in residential sectors. The main aim of proposed
scheme is to decrease the cost of electricity, minimization of peak to average
ration (PAR) and to obtain maximum user comfort. In this scheme, we consid-
ers a single home in which 15 appliances are present. Two Meta heuristic tech-
niques: earth worm algorithm (EWA) and single swarm optimization (SSO)are
proposed and implemented with different price signals. Proposed scheme gives
better results as compared to EWA and SSO.
The rest of the paper is categorized as follows: Sect. 2 defines the related
work. Section 3 discuss the problem statement. In Sect. 4, explains the system
model. Section 5, we discusses the proposed scheme. Computational results are
shown in Sect. 6. Paper findings are presented in Sect. 7 (Table 1).
An Efficient Scheduling of User Appliances 373
Table 1. List of acronyms
DSM Demand side management

GWO Gray wolf optimization
BPSO Binary particle swarm optimization
EMC Energy management controller
DR Demand response
TOU Time of use
IBR Inclind block rate
SG Smart grid
MILP Mixed integer linear programming
GA Genetic algorithm
DP Dynamic programming
CSA Cuckoo Search algorithm
FP Fractional Programming
HSA Harmony search algorithm
EDE Enhance differential evaluation
LOT Length of operational time
PAR Peak to average ratio
RTP Real time pricing
CPP Critical peak pricing
RES Renewable energy sources
AMI Advance metering infrastructure
EDE Enhanced differential evaluation
HEMS Home energy management system
2 Related Work
Mixed integer linear programming (MILP), method is used in [1], to mini-
mize the total electricity bill paid by consumers is the main purpose of this
method. Authors worked on balance load management however, the user com-
fort is neglected in MILP. It shows an exchange between conventional systems
and today’s renewable energy sources (RES). A two-way communication between
utility and consumer through SM, a lot of energy wastage problem covered by
saving this 10–30%. In SG, the big challenge for researchers is cost minimization.
The genetic algorithm (GA) technique used in the paper [2], with the integration
of RES and stored energy a low level of cost minimization achieved. In particular
time changes when electricity price and user demand are higher than the stored
energy is helpful. Authors neglected the deployment and maintenance cost of
storage devices and RES in this technique. One of the big problem is to balance
the load in commercial and residential areas (Table 2).
Table 2. Related work
Technique Achievements Limitations

MILP [1] Reduction in PAR and Comfort preferences are
total electricity cost not considered
GA [2] Cost minimization Deployment and
maintenance cost of
storage devices and RES
are ignored
GA- DSM [3] Electricity consumption Neglected the PAR value
reduced and user comfort
feasibility
MINLP under time Of Cost minimization Neglected the PAR
use (ToU) [4]
DP [5] Cost minimization Installation and
maintenance is ignored
Combination of GA and Cost and peak Ignored comfort
binary partial swarm curtailment preferences and focus
optimization BPSO only residential area
Algorithms [6]
CSA [7] Shifting the load in Neglect the electricity
another time Interval cost
and peak load reduced
BBSA [8] Shifting the load and Consider specific time
electricity price reduced interval in a day and
hardware and software
installation expense
Two point estimation Compute load burden Neglected cost of
method embedded with electricity and PAR
PSO based method [9] value
FP [10] Electricity cost reduction Neglect the PAR and
user comfort
HSA [11] HSA algorithm is Real time
structure, and implementation is not
applications considered
Single knapsack [12] Energy consumption Harder architecture in
optimization considering terms of modeling in real
six layer architecture time scenario
(EDE and HSA) [13] RESs startup and Computational time is
generation cost increased
GWO [16] Solving non-linear The user has to come up
economic load dispatch with ways of handling
problems the constraints
Greedy algorithm [17] Minimized cost and user PAR is ignored
frustration
Though, by using GA-DSM [3] algorithm in maximum hour, 21% of electricity

consumption is reduced in an industry which is very noteworthy. The PAR value
and user comfort feasibility ignored by authors. In [4] authors proposed scheme
of MINLP to solved the cost minimization under the price tariff ToU. Even
though at the peak hour, cost minimization is achieved, however, authors don’t
considered PAR. Cost minimization and the scheduling of gadgets for various
duration achieved by using dynamic programming [5] technique. Authors in [5],
focused only on residential consumption. This achieved by the integration of
RES and ESS’s with SG. Residents are capable of producing the electricity from
RES. An additional electricity could be sold by consumers to their neighbors.
In RES, two important factors are: like installation and maintenance has been
ignored.
Combination of GA and binary partial swarm optimization (BPSO) algo-
rithms proposed in [6]. PAR minimization and electricity cost are the main goal
of this technique. User comfort ignored and it focused only on residential areas.
In DSM, the client can deal with their home appliances by moving the load to
some other time so the load request a key factor in such manner. By shifting the
load and using cuckoo search algorithm (CSA) [7] algorithm, approximately 22%
of peak load has been reduced. The curve of balanced load that is generated by
the CSA algorithm worked on the user partiality for appliance usage, this curve
used then to shift the load.
New binary backtracking search algorithm (BBSA) [8] was proposed for real-
time schedule controller. In comparison to PSO algorithm, home appliances
shifted from peak hour and electricity price reduced 21% per day by using the
load limit.
The two point estimation methods embedded with partial swarm optimiza-
tion (PSO) method is developed by Huang et al. [9] for dropping the computa-
tional complexity in a HEMS. In contrast, this scheme is intelligent enough than
GPSOLHA in the perspective of computational burden. However in HEMS the
cost of electricity and PAR value has not been considered by Author [9]. For
residential appliances, the authors proposed an improved model for HEMS in
[10]. The main goal is to minimize the cost by shifting the appliances. Using the
RTP tariff and DER, fractional programming implemented for HEMS.
The authors discussed the harmony search algorithm (HSA) algorithm in [11].
Authors also defined the searching criteria of different techniques. The primary
steps of HSA is adaptation and used in different fields.
Authors designed a model in [12] for microgrid systems. The microgrid sys-
tems are integrated with the RESs. The main goal of this model is to minimize
the cost of RESs startup and RESs generation cost. The desired objective is
achieved by the combination of EDE and HAS. PAR value is ignored in the
design model.
Authors proposed a model for HEMS with multiple appliances in [13]. Six
Layered are connected with each other to perform better results for the reduc-
tion of PAR and cost. The author in [4] proposed utilizing MINLP method cost
minimization to be accomplished under ToU price signal. The main goal of this
technique is cost minimization. Customers can save maximum energy cost using
MINLP algorithm. Scheduling the load is the core objective of load management
during high demand to low demand time. Evolutionary algorithms are used for
load shifting [15]. All service sides have data sets, where scheduling problem have
been managed to solve the efficiency problem, the industry faced more problem
because of big power consumption appliances. Due to high load users need to use
the energy more intelligently in both residential and commercial sector. Authors
in [15] proposed an algorithm for load management. The main goal of this paper
reduces the electricity cost. All service sides have data sets, where the scheduling
problem have been managed to solve the efficiency problem, the industry faced
more problem because of big power consumption appliances. Customers require
high load in efficient way and more intelligently in the residential and commer-
cial area. Authors proposed a model for cost minimization in [17]. Using an
intelligent decision system (IDSS), minimum cost and minimum PAR problems
were solved. IDSS provides better result communication between the user and
utility. Authors discussed the EDE algorithm in [18]. An updated version of ED
was used. Authors used five trial vectors instead of one. Using three different
random vectors, a new population was created. The mutant and trial vectors
were generated by the fitness function. The authors proposed [19] evolutionary
accretive comfort algorithm (EASA) which is comprised of four postulations.
These postulations are defined according to the time and device bases priori-
ties. Based on the input values EASA generates optimal energy solution pattern
which satisfies the user budget. The author defines three different user budgets
to find the absolute solutions. Ma et al. [20] defines discomfort function for two
different type of gadgets. First category is flexible starting time and the other
is flexible power devices. Authors in [20] considered a multi-objective function
for user comfort and cost parameters. The proposed bat algorithm in [21] can
be applied to obtain the optimum result. By applying this algorithm energy
consumption can be reduced which is simply a non-linear optimization prob-
lem. The important goal of bat algorithm is to decrease the power usage and
increase consumer comfort standards in the residential area. In system models,
mathematical problems and scheduling, the bat algorithm used to solve these
problems.
3 Problem Statement
Total energy consumption, minimum electricity bill, minimum PAR and maxi-
mum user comfort are the most important problems in SG. DSM schemes are
used to overcome the aforementioned problems in the SG. In [1–4] purposed
meta heuristic techniques to solved the energy consumption problem according
to electricity bill and user comfort. Some authors used mathematical solutions
to solve these issues. Some authors [4–6] proposed RES with the integration of
ESS to tackle the energy demand problems. The main aim of all techniques is to
reduce the electricity bill with maximum user comfort. In some hours, when user
demand is higher than the total electricity production. It creates a peak load in
the smart home so HEMS is faced difficulty to maintain the balance. To avoid
such problems, proposed algorithm, which shifts the home appliances from on
peak hours to off peak hours, and thus achieve the minimum cost and maximum
user comfort.
4 System Model
In SG, DSM controller is used for managing the user demand according to the
consumer demand. DSM provides more reliability and proficiency in user tasks.
Different meta heuristic techniques are implemented in DSM to control the user
appliances. DSM techniques manages the load to consume the electricity at off-
peak hours instead on peak hours. Initially, considering single home with 15
appliances. These appliances are categorized into two main categories: schedu-
lable and non-schedulable appliances. First category is further categorized into
two sub types: interruptible and non-interruptible. Non-interruptible appliances
are those appliances that cannot be shifted when they are running and cannot be
switched on as per the user’s requirements. Where interruptible appliances are
the those appliances which can be allocated to various time intervals. Every single
home contains SM. SM decides the operation time of every appliance according
to the power rating. These 15 appliances are taken from [14] details of appli-
ances are shown in Table 3. SM provides the two-way communication between
consumers and service provider. Different price signals (CPP, ToU, RTP) are
used to find the electricity bills. The service provider provides the electricity
price signal. ToU used as pricing unit to calculate the electricity cost. The main
objective of our study is to minimize electricity consumption in order to reduce
the cost and PAR, however, the tradeoff will occur between cost and user com-
fort. Equation 1 is used to calculate the PAR. Cost value is calculated by the
Eq. 2. Total load formula given in Eq. 3. Main architecture of system model is
shown in Fig. 1.
max(loads )
P AR = (1)
avg(loads )

24
App
hour
Cost = (ERate ∗ PRate ) (2)
t=1
app
Load = PRate ∗ App (3)
Table 3. Appliances used in simulation
Appliances Power (kW) Category

Vacuum cleaner 1.2 Interruptible
Sensors 0.01 Interruptible
PHEV 3.5 Interruptible
Dish washer 1 Interruptible
Stove 3 Interruptible
Microwave 1.7 Interruptible
Other occasional loads 1 Interruptible
Clothes washer 1 Non-interruptible
Spin dryer 2.5 Non-interruptible
Oven 5 Base
TV 0.6 Base
PC 0.3 Base
Laptop 0.1 Base
Radio/player 0.2 Base
Coffee maker 0.8 Base
5 Proposed Scheme
Mathematical optimization algorithms try to solve the energy consumption prob-

lems however, with large number of smart devices its harder to present the sat-
isfactory solutions. Behind this problem, use meta heuristic techniques to solve
the energy consumption problem and reduce the cost. The main objective of
EWO algorithm is to reduce the electricity bill and shift the appliances into off
peak hours. SALP algorithm provides the facility to manage the appliances using
different price signals. Different electricity price signals are discussed to define
the cost of electricity for a complete day. In our scheme, we consider RTP tariff.
The RTP is updated for every one hour. Two-way communication requires to
interact with the user for RTP.
5.1 EWO
The reproduction of earthworms states multiple optimization issues, the repro-

duction steps of earthworms are following:
• Each earthworm have the capacity to reproduce off springs and every earth-
worm individual have two kinds of reproduction,
• Every child of earthworm contains all the genetic factor of parents,
• Singular earthworm is moved on next generation, and cannot be changed by
operators.
5.2 SALP
Salps belongs to the family of salpidae. Swarming behavior is one of the most
interesting behavior. In salps, population has two groups: 1. Leader 2. Followers.
The leader is at the front of the chain and rest of the salps are attached behind
it followers. Equation 4 is used to update the position of the leader.

1 Fj + c1 ((ubj − lbj )c2 + lbj ) c3 ≥ 0
xj = (4)
Fj − c1 ((ubj − lbj )c2 + lbj ) c3 < 0
5.3 Updated Population Scheme
Our proposed algorithm provides the facility for consumers to schedule the appli-
ances and reduce the cost while considering the user comfort. The population
size is 30. In our proposed algorithm, the step of reproduction contains two types
of reproductions: Reproduction 1 and Reproduction 2. In our research, we have
implemented a new updated population scheme for scheduling the home appli-
ances. The first step is the initialization of all parameters with the maximum
generation and constant value. A fitness function is defined for choosing the
best solution. After applying the fitness function, two types of reproductions are
applied. The main purpose of the updated population algorithm is to obtain the
maximum solution of appliances. For complexity, mutation and crossover steps
are taken in the proposed algorithm. The major contribution of the proposed
algorithm updates the population according to their fitness function.
Algorithm 1. SALP Algorithm

1: Initialization the salp population xi (i=1,2,......n) considering ub and lb
2: while (end condition in not satisfied ) do
3: Calculate the fitness of each agent (salp)
4: Set Fas the best search agent
5: Update ci value
6: for every salp (xi )) do
7: if (i==1) then
8: Update the position of the leading salp
9: else
10: Update the position of the follower salp
11: end if
12: end for
13: Update the salps
14: end while
15: Return
6 Simulation and Reasoning

In this section, simulation results are briefly discussed. The implementation of
the proposed has been done in MATLAB. The main objective of DSM in the
smart home achieved by proposed scheme. The objective of DSM is maximize
user comfort, minimum electricity cost and minimum PAR. Proposed scheme
results are better as compared to EWO and SALP algorithm. For experimenta-
tion, we considered 15 appliances in the SH. By applying the ToU/CPP price
signal in the smart home, proposed scheme achieved the minimum cost of elec-
tricity and user comfort, however, PAR value is compromised. Figure 1 defines
the electricity cost per hour in cents for the 24 h. It clearly defines the electricity
cost of unscheduled, EWA algorithm, SSO algorithm and updated population
proposed algorithm. Figure 2 shows the total load of aforementioned algorithms
in (kwh), in a day. Figure 3 shows the PAR values of unscheduled, EWA, SSO
and proposed scheme. It clearly shows that the EWA and SSO algorithm outper-
formed the proposed scheme in terms of PAR. EWA and SSO performed better
in terms of PAR as compared to own algorithm. However, there will be a tradeoff
between PAR and user waiting time. The PAR is reduced 49.27%, 50.24% and
42.94% by EWA, SSO and proposed algorithm respectively with the unsched-
uled case. Figure 4 represents the comparison of EWA, SSO algorithm and the
proposed scheme in the regard of electricity cost. In peak hours electricity cost
of appliances increases quickly. To overcome this, scheduling the appliances is
done on, on-peak hours to off-peak hours. The cost of appliances is reduced due
to scheduling. Figure 4 shows the cost values of unscheduled, EWA, SSO and the
proposed algorithm. Proposed algorithm performs better as compared to EWA
and SSO. The user comfort is reduced by 1.75%, 10.23% and 11.76% by the
EWA, SSO and proposed scheme respectively. Considering the total electricity
cost in the form of cents. In Fig. 5, the user waiting time shown. The user wait-
ing time is calculated in terms of user comfort. User waiting time is inversely
proportional to user comfort. By applying the price signal user comfort value
of proposed scheme is low, as compared to EWA and SSO. The user comfort is
reduced by 76.36%, 75.36% and 86.16% by the EWA, SSO and proposed scheme
respectively (Fig. 6).
Fig. 2. Electricity cost
Fig. 3. Load
Fig. 4. PAR
Fig. 5. Total cost
Fig. 6. Waiting time

7 Conclusion
In this paper, proposed algorithm is used for shifting the appliances. The schedul-
ing is based on real-time data of price signal. Proposed scheme results are better
as compared to EWO and SALP algorithm. It is clear that the algorithm intro-
duced works efficiently as compared to EWO and SALP with the parameters of
cost, PAR and user comfort. With the proposed algorithm derived from EWO
and SALP proposed algorithm achieved minimum cost and maximum waiting
time of our proposed scheme. In the future, we will integrate renewable energy
system and ESS with more then one SH in order to minimize cost and maximize
user comfort.
References
1. Bradac, Z., Kaczmarczyk, V., Fiedler, P.: Optimal scheduling of domestic appli-
ances via MILP. Energies 8(1), 217–232 (2014)
2. Mary, G.A., Rajarajeswari, R.: Smart grid cost optimization using genetic algo-
rithm. Int. J. Res. Eng. Technol. 3(07), 282–287 (2014)
3. Bharathi, C., Rekha, D., Vijayakumar, V.: Genetic algorithm based demand side
management for smart grid. Wirel. Pers. Commun. 93(2), 481–502 (2017)
4. Setlhaolo, D., Xia, X., Zhang, J.: Optimal scheduling of household appliances for
demand response. Electr. Power Syst. Res. 116, 24–28 (2014)
5. Samadi, P., Wong, V.W.S., Schober, R.: Load scheduling and power trading in
systems with high penetration of renewable energy resources. IEEE Trans. Smart
Grid 7(4), 1802–1812 (2016)
6. Ullah, I., Javaid, N., Khan, Z.A., Qasim, U., Khan, Z.A., Mehmood, S.A.: An
incentive-based optimal energy consumption scheduling algorithm for residential
users. Procedia Comput. Sci. 52, 851–857 (2015)
7. Cakmak, R., Altas, I.H.: Scheduling of domestic shiftable loads via cuckoo search
optimization algorithm. In: 2016 4th International Istanbul Smart Grid Congress
and Fair (ICSG), pp. 1–4. IEEE, April 2016
8. Ahmed, M.S., Mohamed, A., Khatib, T., Shareef, H., Homod, R.Z., Ali, J.A.: Real
time optimal schedule controller for home energy management system using new
binary backtracking search algorithm. Eng. Build. 138, 215–227 (2017)
9. Huang, Y., Wang, L., Guo, W., Kang, Q., Wu, Q.: Chance constrained optimization
in a home energy management system. IEEE Trans. Smart Grid (2016). https://
doi.org/10.1109/TSG.2016.2550031
10. Nguyen, H.K., Song, J.B., Han, Z.: Distributed demand side management with
energy storage in smart grid. IEEE Trans. Parallel Distrib. Syst. 26, 3346–3357
(2015)
11. Gao, X.Z., Govindasamy, V., Xu, H., Wang, X., Zenger, K.: Harmony search
method: Theory and applications. Comput. Intell. Neurosci. 2015, 39 (2015).
https://doi.org/10.1155/2015/258491
12. Mahmood, A., Baig, F., Alrajeh, N., Qasim, U., Khan, Z.A., Javaid, N.: An
enhanced system architecture for optimized demand side management smart grid.
Appl. Sci. 6, 122 (2016). https://doi.org/10.3390/app6050122
13. Zhang, J., Wu, Y., Guo, Y., Wang, B., Wang, H., Liu, H.: A hybrid harmony
search algorithm with differential evolution for day-ahead scheduling problem of a
microgrid with consideration of power flow constraints. Appl. Energy 183, 791–804
(2016)
14. Shirazi, E., Zakariazadeh, A., Jadid, S.: Optimal joint scheduling of electrical and
thermal appliances in a smart home environment. Energy Convers. Manag. 106,
181–193 (2015)
15. Logenthiran, T., Srinivasan, D., Shun, T.Z.: Demand side management in smart
grid using heuristic optimization. IEEE Trans. Smart Grid 3(3), 1244–1252 (2012)
16. Yuce, B., Rezgui, Y., Mourshed, M.: ANN-GA smart appliance scheduling for
optimised energy management in the domestic sector. Energy Build. 111, 311–325
(2016)
17. Ogwumike, C., Short, M., Abugchem, F.: Heuristic optimization of consumer elec-
tricity costs using a generic cost model. Energies 9, 6 (2016)
18. Islam, S.M., Das, S., Ghosh, S., Roy, S., Suganthan, P.N.: An adaptive differential
evolution algorithm with novel mutation and crossover strategies for global numer-
ical optimization. IEEE Trans. Syst. Man Cybern. Part B (Cybern.) 42, 482–500
(2012)
19. Khan, A., Javaid, N., Khan, M.I.: Time and device based priority induced comfort
management in smart home within the consumer budget limitation
20. Ma, K., Yao, T., Yang, J., Guan, X.: Residential power scheduling for demand
response in smart grid. Int. J. Electr. Power Energy Syst. 78, 320–325 (2016)
21. Mohan, N., Sivaraj, R., Pyira, R.D.: A comprehensive reviews of bat algorithm and
its application to various optimization problems. Asian J. Res. Soc. Sci. Humanti.
6, 676–690 (2016)
Pro Utility Pro Consumer Comfort
Demand Side Management in Smart Grid
Waleed Ahmad1 , Nadeem Javaid1(B) , Basit Karim2 , Syed Qasim Jan3 ,

Muhammad Ali4 , Raza Abid Abbasi1 , and Sajjad Khan1
1
COMSATS University, Islamabad 44000, Pakistan
2
COMSATS University Islamabad, Abbotabad, Pakistan
3
University of Engineering and Technology, Peshawar, Pakistan
4
King Saud University, Riyadh, Saudi Arabia
Abstract. Now a days, energy is the essential resource and due to

increase in power demand, traditional resources are not enough to ful-
fill the requirement of todays need. The researchers are working on new
approaches to enhance and improve the power load demand. The increas-
ing demand of electricity creates peaks on utility. Therefore an improved
Home Energy Management System (HEMS) is necessary for the automa-
tion of smart home to reduce the cost and peaks on utility. In this paper
work, our objective is pro utility and pro-consumer comfort which means,
the decrease in Peak to Average Ratio (PAR) in order to reduce the
stress on the utility while increasing user comfort. In this Research, we
have proposed a new technique called Random Cell Elimination Scheme
(RCES) with Demand Side Management (DSM) for a home appliance
scheduling. To make the system more effective, we have utilized two pric-
ing systems: Time of Use (ToU) and Real Time Pricing (RTP) in our
experiment. The simulation results are compared with two heuristic opti-
mization schemes: Bacterial Foraging (BFA) and Firefly Algorithm (FA).
The experimental results shows that the proposed scheme performed 80%
better than BFA and FA in reducing PAR and user discomfort.
1 Introduction
The smart grid is more powerful than the traditional grid system because of bidi-
rectional communication between the source and end users which makes it more
scalable and efficient [1]. The Demand Response (DR) helps in the interaction
between the power source and consumers to provide more efficient and reliable
system [2]. The main role of DR is to reduce the burden on the power source
by motivating the consumers to change their energy usage behavior according
to price tariff provided by a utility. It has been analyzed that more than 40% of
energy is consumed in a residential area around the world [3].
The international energy outlook in 2013 assumes that by 2040, it will rise to
more than 56% [4]. To tackle all these challenges, smart grid emerged with the
https://doi.org/10.1007/978-3-030-15035-8_36
386 W. Ahmad et al.
concept of more reliability, productivity, and cost-effectiveness with the most

important feature like dynamic control. Demand Side Management (DSM) in
smart grid plays an important role by utilizing optimization techniques in an
adaptable way to sustain the grids balance [5]. The operations of DSM with
Home Energy Management System (HEMS) helps the consumer to control the
load demand dynamically. Fundamentally, the purpose of DSM is to educate the
electricity users to change their usage pattern in order to reduce the cost and bur-
den on the power grid [6]. The DSM strategies include control and management
of load demand and appliance scheduling. However, despite the above-mentioned
strategies, load shifting is used vastly in past research work to manage the load
by DR [7]. This helps the consumers to transfer the load from ON peak to Off-
peak hours to provide incentive to the consumers while lowering the stress on
the main grid to reduce the Peak to Average Ratio (PAR) [8].
The main purpose of the above-discussed approaches is to aware the users to
reduce the PAR and load. Different type of scheduling algorithms is adopted by
DSM to optimize the load demand pattern [9]. As we know, the residential sector
has many household appliances which are running time to time in a whole day,
which attracts the attention of researchers. To overcome this problem, the HEMS
[10] is proposed with the different type of pricing tariffs. HEMS is responsible
to handle and monitor the energy demand in a smart home dynamically [11].
The real-time dynamic scheduling is a difficult task and done by using dynamic
programming [12].
The proposed work of this paper is presented as Follows:
Random Cell Elimination Scheme (RCES): In this paper, our objective is
to present a new heuristic optimization scheme. we analyze the existing heuristic
algorithms by their performance and nature and after evaluating the Bacterial
Forgery Algorithm (BFA) and FireFly Algorithm (FA), we determine that there
need to a more efficient system which can improve multiple limitations such that
reducing the PAR and user discomfort.
2 Related Work
The utilization of smart grid elements by deploying DSM and DR is a demanding

task. In the last few years, many optimization schemes have been proposed and
experimented considering single or multi goals with multiple trade-off objectives.
Some systematic approaches are presented in Table 1. In many articles [10,11],
the main focus of research was to reduce the user discomfort, lessen the cost
on the user side and reduction in PAR to benefit the utility directly or by indi-
rectly using some heuristic algorithms, mathematical problems, and statistical
schemes. These types of techniques [12–14], cut down the stress of the power by
transferring the load from ON-Peaks to OFF-Peaks time intervals to reduce the
cost while giving the power to the user to operate a different kind of appliances
according to their preference and comfort. The researchers in [14,15], compose
the problems for multi-objective models, wherein [16], express the problem as a
Pro Utility Pro Consumer 387
single objective optimization. This system schedule the smart devices of an intel-
ligent home. Furthermore, the objective was to stable the load by a mixed integer
linear programming to reduce the cost and PAR. The simulation conducted in
the Czech Republic by taking real-time pricing.
In [14,17], the article focus was to reduce the total cost and reduce the
electricity bill of an individual. Simulations in [18], performed on real-time data
received from Energy and Climate Center UK by taking real time tariffs with
multi-objective mixed integer programming. The high demand determination
scheme is proposed in [15], for residential area using Real Time Pricing (RTP)
for multiple cases. The finite smart appliances are taken with quasi-random
methods and the recursive equation is formed to check the high load demand.
Intelligent HEMS model is presented in [19]. In this model, oU pricing tariff
is used to reduce the total cost without affecting the consumer satisfaction.
In this concern [20], a cost-effective greedy iterative distributed model is used
for multiple consumers to overcome the cost problem in a grid. The simulation
outcomes expose that electricity cost is degraded and PAR in decreased. The
distributed generation model is proposed in [21], which effectively reduced the
cost and waiting time by allowing bidirectional electricity flow. Results show
that for all cases, the 32.73% of energy cost is saved. The user discomfort is
discussed in [22]. In this research paper, the demand response issue is developed
in HEMS with a support of the Markov decision model. The DSM controller is
employed with Genetic Algorithm (GA), Binary Particle Swarm Optimization
(BPSO) and Colony Optimization Algorithm (COA) in [23]. In this scheme, GA
outperforms the rest of the schemes. The cost minimization and waiting time is
directly related to user comfort as discussed in [24].
3 Problem Statement
Traditional grid system has many flaws and limitations. From the last decade,
smart grid attracts the attention of the researcher towards its benefits and advan-
tages. The smart grid contains two-way communication between consumer and
power source. This absolute goal of the smart grid to generate enhancement
within supply, to organize and regulate the whole process of electrical energy
[25]. To meet the requirement of consumers for energy demand by utilizing dif-
ferent techniques identified by DSM. As we know, a single home has too many
electrical appliances which are running time to time with the different type of
power ratings. With the growth in population, power need in the residential side
is raising gradually [26]. By means of employing DSM, a person can help to
eliminate the price as well as handle the unwanted using electrical power by way
of arrangement appliances properly. Through DSM, scheduling of smart devices
is done by knowing the peak hour and estimating the price through the present
hour. These kinds of switching lesson the price but load the utility and therefore
associated with general shortage of energy.
388 W. Ahmad et al.
Table 1. Summarized related work
Technique(s) Achievement(s) Limitation(s)

Linear Programming The proposed scheme Only focuses on cost
reduce cost and PAR and PAR, user comfort
is ignored
PSO GA and MKP This model reduces cost UC is ignored
and PAR
Mixed linear and integer Cost and Peak load is User discomfort
linear programming reduced increased
Multi-objective mixed Generation cost is PAR ignored
linear programming reduced while taking
user comfort
Double co-operative Cost is reduced on both Smart vehicle charge is
game theory user and utility side avoided
Game theocratic energy Reduce the load on User comfort is not
management utility and user bill considered
Quasi random process Load management is PAR and user comfort
done with recursive is ignored which burden
formula the utility
Greedy iterative Cost, PAR reduction Stress the utility,
algorithm with load shifting increase the energy cost
FA Increase the user Cost is high
comfort and reduce the
PAR
Reinforcement learning User satisfaction with PAR is high
reduction in cost
BFA, BPSO and ACO Cost and PAR is Renewable energy
minimized with resources installation is
improvement in comfort not considered
WDO Waiting time is reduced Installation cost is not
observed with energy
cost
4 Proposed Solution
The main focus of our work is to reduce the peak load and total cost which is
calculated in Eq. 1. The RTP and ToU 24-h onward day pricing is used in the
simulation. The central challenging part is load shifting from On-Peak hours to
Off-peak hours in order to tackle to load balance and reduce cost. The load shift-
ing is done based on the fitness criterion which is presented in Eq. 2. Now in order
to determine the total load, the Eq. 3 is employed, where alpha represents the
status of appliances in 0’s and 1’s. Table 1 shows the appliances and their power
rating according to their classifications of interruptible and non-interruptible.

24
App
hour
Cost = (ERate ∗ PRate ) (1)
hour=1
⎧ ieN p
⎨lod ≥ mean(LU s hour
od ), ERate ≤ mean(ERate )
Np ieN p
Ff = min l > (std(Lod )) ∧ lod < mean(LU
U s s
od )
(2)
⎩ od hour
ERate > mean(ERate )
App
Lod = PRate ∗ App (3)
Our main objective is not only lessened the Cost as defined in Eq. 5 but also
decrement in PAR by applying Eq. 6. Along with cost and PAR reduction, we
also have objective to achieve efficient load shifting in DSM as assessed in Eq. 7.
Ojb1 = min(Cost) (4)
Ojb2 = min(P AR) (5)
Ojb3 = min(Load) (6)
To achieve this goal, we have to split a day into 24-h time slots. Furthermore,
these time slots are split into two parts (high price hours and low price hours).
Now the main challenge is to transfer the high load to low load hours according
to their power ratings classification. This approach will effectively decrease the
cost and PAR due to load balance. For calculation of PAR, we have used Eq. 8.
max(LSod )
P AR = (7)
Average(LSod )
5 System Model
In our proposed work, simulation is performed on a smart home which consists
of fifteen different appliances that need to be scheduled. Smart appliance Length
of Operation (LoT) and power information is collected from end users. The pro-
posed system model is divided into three sections such as DSM, Supply Side
Management (SSM), and Communication unit. SSM have all the information
related to energy generation and pricing tariff. Energy Management Controller
(EMC) is used by DSM for appliance scheduler. The objective of handling the
load is to bound the consumers from excessive usage of electricity to lower the
burden on utility. The communication unit is responsible for the transfer of
the pricing scheme and dynamic demand-response data between DSM and SSM.
Home Area Network (HAN) is used in a smart home for communication between
the smart appliances and EMC through Wi-Fi and Zig-Bee protocols. Further-
more, there are three classifications of smart devices according to the schedule:
the Interruptible, Un-interruptible, and base load. In our paper, three meta-
heuristic scheme are selected, including FA, BFA and RCES. These schemes are
adopted to manage the smart appliance in HEMS. The central idea of this sys-
tem is to decrease the consumption throughout the peak time to diminish the
cost and PAR. Simulations is performed on actual presumed smart appliances
that are usually split by a few distinct types like: Interruptible, Non-interruptible
and base load (Fig. 1) (Table 2).
390 W. Ahmad et al.
Table 2. Appliances used in simulations
Appliances Power rating (kwh) LOT (h)

Refrigerator 1.666 24
Vacuum cleaner 0.7 0.7
Water pump 1 8
Washing machine 1.4 3
Cloth dryer 5 4
Dishwasher 1.32 3
Water heater 5 8
Iron 2.4 3
AC 1.5 8
Cooker 0.225 4
Toaster 0.8 1
Printer 0.011 2
Light 0.18 12
Blender 0.3 2
Oven 2.4 4
6 Meta-heuristic Algorithms
Several optimization methods have already been deployed to handle the smart
grid problems. In this regard, several effective heuristic techniques have already
been utilized to produce a strong strategy to tackle energy and smart grid prob-
lems. In that study, we have applied two meta-heuristic systems, BFA and FA.
6.1 BFA
Nature neglects those animals which have bad foraging methods and support
those who have surviving strategies. Initially, BFA algorithm was developed
by [27]. After many productions, poor versions usually are changed or even
re-designed using healthy versions. The stochastic character of the algorithm

allows the cell to improve statistically and collaboratively approaching the most
effective solution.
6.2 FFA
FFA merge as an obstacle solver [28]. This algorithm helps in giving optimized
solutions to many different optimizations obstacles. Firefly is known as multi-
handler due to its multi-model feature. In our research, we have adopted a new
scheme to tackle the scheduling index.
6.3 RCES
In our research, we have implemented a new Random Cell Elimination Scheme
(RCES). In RCES, random data is generated and the distance of the closest cell
is evaluated. The population is assessed by the number of steps in the defined
position and fitness function is defined to cut back the overall cost. The pro-
duction of the Cell is evaluated and replaced with the fittest position. The new
generation of cell is produced in next step with random population. This new
generation eliminates the old cells based on objective and fitness evaluation. The
main purpose of RCES is o obtain the optimum solution of appliance scheduling
in smart home. However, optimized home scheduling task is a difficult goal to
achieve. Simulation results demonstrate that our system successfully lower the
price and PAR as trade-off between price and delay time, so in order to achieve
one factor, other might be sacrificed.
1. Initialization of the population
2. Fitness Evaluation of each Cell
3. Reproduction
4. Replace previous place with good one.
Algorithm 1. Algorithm for RCES

1: Initialize (P oP, Np , Ne , Nr , Nc , Ci )
2: Evaluate the initial;
3: for t = 1:Ne do
4: for t = 1:Nr do
5: for t = 1:Nc do
6: for t = 1:Np do
7: end For
8: end For
9: end For
10: Calculate Best Solution for task;
11: Pbest=intensity.best with lowest time;
12: Allocate Solution;
13: Evaluate New Solution;
14: end For
15: Find the health(fitness);
16: Select the gbest;
392 W. Ahmad et al.

In the simulation section, we have selected two heuristic algorithm and evaluation
is done on the foundation of Price, PAR and delay time. Two pricing tariffs ToU
and RTP are utilized in the simulation. The simulation results are evaluated on
two pricing schemes RTP and ToU by using proposed heuristic scheme RCES.
We have used fifteen appliances of smart home that need to be managed and
controlled.
The total price, load per hour and the delay time is denoted in words of cost
and respective time slot. Figures 2 and 3 represents the total electricity cost in
cents and Kwh of the single smart home by using RTP and ToU price tariff in
term of 24-h time schedule. simulation results shown that each scheme is out-
performing the unscheduled case. The total cost in term of KWh in represented
for 24-h time scheduled. Simulation result reveal that heuristic schemes perform-
ing better than unscheduled. The proposed scheme efficiently reducing the cost
during peak hours as shown in graph.
In Figs. 4 and 5, The total load in term of 24 h time interval is shown which
clearly indicate that during peak hour, every heuristic scheme is beating the
unscheduled case. The peak load of unscheduled case is higher while on the
other side heuristic algorithms reduce considerably load in peak times.
Fig. 2. Electricity cost (Cents) TOU
Fig. 3. Electricity cost (Cents) RTP

Fig. 4. Load (KWh) TOU
Fig. 5. Load (KWh) RTP
The main objective of our approach is to reduce the PAR to lower the burden
on the utility and reduce waiting time. PAR explain the power consumption pat-
tern of the user load and effects the power grids. In Figs. 6 and 7, unscheduled and
scheduled PAR is illustrated. The simulation results shows that presented app-
roach RCES reduce considerably amount of PAR as compared to BFA and FA,
which shows that proposed approach is better than other two heuristic schemes.
However, PAR is related to burden on power grids. Simulation results reveal all
the heuristic schemes minimize the PAR. However our proposed heuristic opti-
mization scheme reduce PAR to 80% and beats the other schemes. So we can say
that our new scheme is pro utility and pro user comfort achievement in smart
grid.
Our New proposed scheme also shows the effectiveness in reducing total cost
as compared to unscheduled scenario. All heuristic algorithm shows effective-
ness in reducing the cost. Reducing cost and PAR using heuristic algorithms is
challenging objective because of trade off between cost and delay time. Figures 8
and 9, the results represent the total energy cost for BFA, FA and RCES using
ToU and RTP pricing tariffs.
Figures 10 and 11, present the waiting time which is mainly concern to user
comfort. User comfort is achieved when a consumer faces less waiting time in
operation of electrical appliances. The results shows that our approach is a pro
394 W. Ahmad et al.
Fig. 6. PAR (ToU)
Fig. 7. PAR (RTP)
Fig. 8. Total cost (ToU)
user comfort because of increase in the user comfort by reducing the waiting
time to 50% which is a major contribution of our work.
In order to analyze the model, the confidence level is measured in term of PAR
and waiting time. Simulation results indicates less gap even in random values.
BFA is an old optimization technique which mainly focuses on production and
Fig. 9. Total cost (RTP)
Fig. 10. Waiting time (ToU)
Fig. 11. Waiting time (RTP)
living of bacteria, while Firefly has known as problem solver because they work
on flashing behavior and distance of the fittest between them. The results reveal
that RCES is efficiently reduce the cost and PAR while scheduling the appliances
towards On-Peak to Off-Peak hour without any stress on the Power Grid.
396 W. Ahmad et al.
8 Conclusion
DSM is used for the control and management of electrical appliances used in
a smart home in order to balance the load demand. In most cases the energy
load is shift to new renewable sources to ease the user but it is not sufficient.
To tackle the above mention problem an optimization technique is adopted. The
new scheme RCES has been proposed which efficiently reduce PAR and user
discomfort. The experiments results clearly revealed that our propose scheme is
better than BFA and FA. The high load is shifted to low peak hours in order
to balance the load and reduce the stress on utility. It is observed that there is
always a tradeoff between cost, PAR and user comfort. The results show that
fitness function improve the performance. The experiment is conducted on the
day ahead ToU and RTP pricing schemes which originally examine the behavior
of optimization schemes. To check the performance of the ToU and RTP pricing
schemes, price forecasting is an essential feature.
References
1. Ipakchi, A., Albuyeh, F.: Grid of the future. IEEE Power Energy Mag. 7(2), 52–62
(2009)
2. Medina, J., Muller, N., Roytelman, I.: Demand response and distribution grid
operations: opportunities and challenges. IEEE Trans. Smart Grid 1(2), 193–198
(2010)
3. Gul, M.S., Patidar, S.: Understanding the energy consumption and occupancy of
a multi-purpose academic building. Energy Build. 87, 155–165 (2015)
4. Khalid, A., et al.: Towards dynamic coordination among home appliances using
multi-objective energy optimization for demand side management in smart build-
ings. IEEE Access 6, 19509–19529 (2018)
5. Gellings, C.W.: The concept of demand-side management for electric utilities. Proc.
IEEE 73(10), 1468–1470 (1985)
6. Broehl, J.H., Huss, W.R., Skelton, J.C., Sullivan, R.W., Limaye, D.R., Sloane,
B.D., Davis, T.D., Blevins, R.P.: Demand-side management, vol. 1. Overview of
key issues. Final report. No. EPRI-EA/EM-3597-Vol. 1. Battelle Columbus Labs.,
OH (USA); Synergic Resources Corp., Bala-Cynwyd, PA (USA) (1984)
7. Davito, B., Tai, H., Uhlaner, R.: The smart grid and the promise of demand-side
management. McKinsey Smart Grid 3, 8–44 (2010)
8. Gelazanskas, L., Gamage, K.A.: Demand side management in smart grid: a review
and proposals for future direction. Sustain. Cities Soc. 11, 22–30 (2014)
10. Shirazi, E., Jadid, S.: Optimal residential appliance scheduling under dynamic
pricing scheme via HEMDAS. Energy Build. 93, 40–49 (2015)
11. Matayoshi, H., Uehara, T., Ludin, G.A., Sharma, A., Datta, M., Senjyu, T.: Unin-
terruptible smart house equipped with a single-phase dq-transformation system.
J. Renew. Sustain. Energy 8(2), 025101 (2016)
12. Mesari, P., Krajcar, S.: Home demand side management integrated with electric
vehicles and renewable energy sources. Energy Build. 108, 1–9 (2015)
13. Adika, C.O., Wang, L.: Smart charging and appliance scheduling approaches to
demand side management. Int. J. Electr. Power Energy Syst. 57, 232–240 (2014)
14. Gao, B., Zhang, W., Tang, Y., Mingjin, H., Zhu, M., Zhan, H.: Game-theoretic
energy management for residential users with dischargeable plug-in electric vehi-
cles. Energies 7(11), 7499–7518 (2014)
15. Vardakas, J.S., Zorba, N., Verikoukis, C.V.: Power demand control scenarios for
smart grid applications with finite number of appliances. Appl. Energy 162, 83–98
(2016)
16. Zhang, Y., Zeng, P., Li, S., Zang, C., Li, H.: A novel multiobjective optimization
algorithm for home energy management system in smart grid. Math. Probl. Eng.
2015, 19 (2015)
17. Jovanovic, R., Bousselham, A., Bayram, I.S.: Residential demand response schedul-
ing with consideration of consumer preferences. Appl. Sci. 6(1), 16 (2016)
18. Qdr, Q.: Benefits of demand response in electricity markets and recommendations
for achieving them. US Department of Energy, Washington, DC, USA, Technical
report (2006)
19. Abushnaf, J., Rassau, A., Grnisiewicz, W.: Impact on electricity use of introducing
time of use pricing to a multiuser home energy management system. Int. Trans.
Electr. Energy Syst. 26(5), 993–1005 (2016)
20. Chavali, P., Yang, P., Nehorai, A.: A distributed algorithm of appliance schedul-
ing for home energy management system. IEEE Trans. Smart Grid 5(1), 282–290
(2014)
21. Elkazaz, M.H., Hoballah, A., Azmy, A.M.: Artificial intelligent based optimization
of automated home energy management systems. Int. Trans. Electr. Energy Syst.
26(9), 2038–2056 (2016)
22. Wen, Z., O’Neill, D., Maei, H.: Optimal demand response using device-based rein-
forcement learning. IEEE Trans. Smart Grid 6(5), 2312–2324 (2015)
23. Rahim, S., Javaid, N., Ahmad, A., Khan, S.A., Khan, Z.A., Alrajeh, N., Qasim, U.:
Exploiting heuristic algorithms to efficiently utilize energy management controllers
with renewable energy sources. Energy Build. 129, 452–470 (2016)
24. Rasheed, M.B., Awais, M., Javaid, N., Iqbal, Z., Khurshid, A., Chaudhry, F.A.,
Ilahi, F.: An energy efficient residential load management system for multi-class
appliances in smart homes. In: 2015 18th International Conference on Network-
Based Information Systems (NBiS), pp. 53–57. IEEE (2015)
25. California Energy Commission. Energy Efficiency and Conservation: Trends and
Policy Issues. No. 100-03. Report, May 2003
26. Jacob, L., Jeyakrishanan, V., Sengottuvelan, P.: Resource scheduling in cloud using
bacterial foraging optimization algorithm. Int. J. Comput. Appl. 92(1), 14–20
(2014)
27. Passino, K.M.: Biomimicry of bacterial foraging for distributed optimization and
control. IEEE Control Syst. 22(3), 52–67 (2002)
28. Fister, Jr., I., Yang, X.-S., Fister, I., Brest, J., Fister, D.: A brief review of nature-
inspired algorithms for optimization. arXiv preprint arXiv:1307.4186 (2013)
Efficient Scheduling of Smart Home
Appliances for Energy Management
by Cost and PAR Optimization
Algorithm in Smart Grid
Sahibzada Muhammad Shuja1 , Nadeem Javaid1(B) , Sajjad Khan1 ,

Hina Akmal2 , Murtaza Hanif3 , Qazi Fazalullah1 , and Zain Ahmad Khan4
1
2
University of Lahore (Islamabad Campus), Islamabad 44000, Pakistan
3
Central South University, Changsha 410083, China
4
COMSATS University Islamabad, Abbottabad 22010, Pakistan
Abstract. As the energy demand for consumption is comparably higher

than the generation of energy, which produce the shortage of energy.
Many new schemes are being developed to fulfill the energy consumer
demand. In this paper, we proposed our meta-heuristic algorithm Runner
Updation Optimization Algorithm (RUOA) to schedule the consumption
pattern of residential appliances. We compared the results of our scheme
with other meta-heuristic algorithm Strawberry Algorithm (SBA) and
Firefly Algorithm (FA). Critical Peak Price (CPP) and Real Time Price
(RTP) are the two electricity pricing scheme that we used in this paper
for calculation of electricity cost. The main objective of this paper is to
minimize the electricity cost and Peak to Average Ratio (PAR). However,
consumer comfort is not satisfied.
Keywords: SA · FA · RUOA · Meta-heuristic techniques ·

Home Energy Management System · Smart Grid
1 Introduction
The traditional Grid (TG) has insufficient capabilities to solve electricity grid chal-
lenges: security of transmission line, bi-directional communication scalability and
robustness against any fault [1]. Therefore, advanced architecture of the TG is
strongly wished to overcome these challenges in an efficient way. Smart Grid (SG)
integrates various communications information to TG. SG allow consumer to con-
trol electricity consumption through bi-directional communication between con-
sumers and source via Smart Meter (SM) in Advanced Metering Infrastructure
(AMI) [2]. Demand Side Management (DSM) is the main section of SG to cre-
ate balance for both demand and supply sides [3]. DSM has two main features:
https://doi.org/10.1007/978-3-030-15035-8_37
Efficient Scheduling of Smart Home Appliances 399
Demand Response (DR) action taken by consumer on dynamic pricing scheme and
load management to schedule the electricity consumption in an efficient way. Elec-
tricity consumption is optimize to reduce electricity cost by DR Program in [4] i.e.
different electricity consumption patter dynamic electricity price.
The main concern of SG is to reduce the PAR, electricity bills and maximize
the consumer comfort. Electricity bills and PAR are reduced at demand side by
efficient management of electricity consumption pattern. Utility provide differ-
ent pricing scheme: Real Time Price (RTP), Critical Peak Price (CPP), Inclined
Block Rate (IBR) etc. Electricity can be managed for the home appliances for
reduction of cost and Peak to Average Ratio (PAR) by pricing schemes. Cur-
rently many DSM mechanisms deployed to overcome previously mention chal-
lenges. In past research work, many of non-heuristic technique are presented to
optimize the home appliances [5], the researcher used Integer Linear Program-
ming (ILP) for reducing load at low price time slot to increase consumer comfort
(CC). However, this technique is not consider for complex energy consumption
pattern.
Scheduling of residential appliances in architecture of SG is provided by Home
Energy Management System (HEMS). HEMS is an intelligent system that opti-
mizes the load consumption pattern in peak hours. In previous work of HEMS,
authors categories the load management as real-time and predictable based. In
[6], predictable approach of load management is deployed. However, this app-
roach is costly and complex to produce uncertainty of solution. While real-time
optimization is employed to overcome the uncertainty issues that is deployable
for large scale of area [7].
In [8], real-time approach of load management is deployed for peak hours.
The exclusively management of energy utilization pattern in residential area is
feasible for optimization of different appliances at home. However, CC is not
satisfied. In this paper, we deployed the meta-heuristic algorithm for real-time
environment. In our work, Strawberry Algorithm (SA) and Firefly Algorithm
(FA) are simulated for their results to compare with our proposed scheme Runner
Updation Optimization Algorithm (RUOA). We developed a new scheme to get
optimal solution for reduction of total electricity cost and PAR as compared
with SA and FA scheme, while CC is sacrificed in our scheme. However, there
is trade-off exist for CC to cost and PAR. Moreover, many spaces exist to fill
the area of load management and more researches are going on to overcome the
energy crises.
The rest of the work is organized as follows: In Sect. 2, we present the related
work for scheduling of appliances in according with different technique. Section 3
explains the problem statement to tackle the issues. In Sect. 4, we discussed
our proposed scheme in detail along with pricing scheme that has been used.
Scheduling techniques are presented in Sect. 5 to work out our proposed scheme.
Simulation results are discussed in Sect. 6. At the end paper conclusions are
drawn in Sect. 7.
400 S. M. Shuja et al.
2 Related Work
In HEMS lot of research has been presented by many authors for optimization
of residential appliances’ scheduling. Researchers trying to induce the efficient
scheduling of electricity consumption pattern by the appliances placed at any
home. Different authors have main objective to consumed electricity in reliable
and efficient way by reducing electricity cost and PAR or increasing CC. Further,
some related research works are discussed below and summarized in Table 1.
Authors highlighted the multi residential electricity load scheduling problem
in [9]. According to author, in previous research only single residential area is
considered, to keep this thing in mind they want to maximize the user comfort for
large residential area. They proposed electricity load scheduling algorithm named
as PL-Generalized Benders Algorithm. The aim of their proposed algorithm is
to schedule the load of appliances and maximize the user comfort. However, it
is not defined how will they schedule load when appliances are near optimal to
each other.
Discussion about the trend by authors in [10], to schedule the electrical load
from on peak hours to off peak hours. The consumer needs scheduling in an
online manner that they can easily find out the prices of electricity and schedule
the load of their appliances according to that manner. To make electricity users
aware about their load scheduling of appliances online Load Scheduling Learning
(LSL) algorithm is proposed in [10]. The aim of their algorithm is to reduce PAR.
However, cost of electricity and performance of proposed algorithm is not defined
by them.
Community based cooperative energy scheme is proposed in [11]. The aim of
proposed scheme is to discuss the electricity cost consumption of the user and
smart grid. In their work, they want to minimize the cost and PAR, to fulfill
their aim they considered community between SG to consumer and consumer to
SG. To evaluate the results of their proposed C2C (community to community)
scheme, MATLAB is used. The cost is minimized because of reducing PAR dur-
ing on peak hours. However, User comfort and security issues are compromised.
In [12], the HEMS system is proposed to schedule the operation of electric
appliances. Authors perceive about Quality of Experience in their work because
they want to check the effects of the proposed system. For this purpose they
consider the profiles of the users and also proposed allocation algorithm. By
applying the proposed system and algorithm electricity cost is minimized. On
the other hand, they have not consider PAR and CC in their proposed work.
The Binary Particle Swarm Optimization (BPSO), Genetic Algorithm (GA)
and cuckoo search meta-heuristic algorithm is used in [13]. In their paper, the aim
of authors is to schedule the load of homes appliances and convert them in smart
homes. They also want to reduce the peak load and electricity bill reduction.
To schedule the appliances load of the homes they also used renewable energy
resources in their work. The electricity bill is minimized. However, the electricity
bill can be reduced further.
The multiple problems such as reducing electricity bill cost and peak reduc-
tion is considered in [14]. They want to solve above mentioned problems, to
solve these problems meta-heuristic algorithm is proposed in their work. The

MATLAB is used to perform the simulations of their proposed meta-heuristic
algorithm. The simulations show the reduction of peak during peak hours and
it also affects the electricity bill. On the other hand, the CC is compromised.
The GA and Grey Wolf Optimization (GWO) algorithms are proposed in
[15]. The aim of these proposed algorithms in this work is to minimize the peak
load of SHs. They also want to reduce the electricity bills of the users without
compromising their comfort. Multiple SHs are considered at different time slot
to check the results of their proposed algorithms, which shows that reduction in
the load and electric bill is performed. However, the performance of optimization
is decreased.
Short term decision making model is proposed in [16]. In this paper, the renew-
able energy resources are used to manage the load of SHs and minimize the elec-
tricity cost. The mixed-ILP method is also used in this paper to make system more
efficient. The proposed model with this method also gives the offer for electricity
to use in fewer rates at specific hours. However, the implementation cost of this
proposed model on the basis of renewable energy resources is very high.
3 Problem Statement
The reduction of electricity cost and energy management in SG is our main

objective due to irregular behavior of energy consumption. Consumer comfort
is usually neglected while considering reduction of electricity price. A scheme is
presented in [17], to optimize the operation of residential appliances in such a
way that minimize the total electricity consumption cost and maximize the CC.
There is always trade off exist between cost and CC.
A model of Energy Management System (EMS) is presented in [18]. It pro-
poses a meta-heuristic technique with RTP and IBR pricing signal for efficient
optimization of home energy to manage the home appliances. Simulations depict
that the significant minimization of electricity cost and PAR. However, the CC
is not considered. In this paper, we performed simulation on SBA and FA algo-
rithms to compared the results with our new proposed RUOA algorithm for
scheduling of home appliances.
3.1 System Description
The proposed system model is shown in Fig. 1. Our system model is working for
scheduling SH appliances. Cost of electricity and PAR is increasing periodically
so there is need of well adaptable system, which can minimize average electricity
price and PAR. A HEMS is used for optimization of home appliances to regulate
the schedule of their operational time. HEMS ease the consumer with respect
of reduction in electricity dissipation and electricity cost. For optimization the
appliances are classified into three different categories; Shiftable appliances, Non-
Shiftable appliances and fixed appliances. Shiftable are those appliances whose
operational time can be shifted from peak hour time slots to non-peak hour time
slots, while Non-Shiftable operational time cannot be shifted or interrupted dur-

ing their operational time and fixed appliances operate on the basis of consumer
demand which cannot be fixed for specific time slots. Appliances of different
classification with their power rating are given in Table 2.
Fig. 1. Proposed system model
In SG system consumer can bi-directionally communicate with utility

through SM, this communication can give information to the consumer for
recently consumed electricity and estimated cost of electricity. In this paper
we used CPP and RTP pricing scheme through which load is shifted from peak
hours to non-peak hours, which enable us to know the energy consumed and cost
charged against per hour. We can reduce our electricity cost by knowing the con-
sumption of energy. In our work, we used our proposed algorithm RUOA in EMC
of SM, compared the result with SA and Firefly algorithm. The results shows
that proposed algorithm RUOA perform in a better way from other algorithm
for scheduling of residential appliances.
To estimate the cost of electricity CPP and RTP pricing scheme is used in
our work. The main objective of our work is to reduce the consumption of energy
to reduce PAR and electricity cost. In Eq. 1, total electricity cost is estimated
with PAR reduction.

24
App
Hour
Cost = (ERate × PRate × App(hour)), App(hour) = [1/0] (1)
hour=1
Hour App
In above equation ERate shows the electricity cost per hour, PRate shows the
power rating of each appliance and App(hour) values present the ON/OFF status
of appliances. In Eqs. 3 and 4 total load and PAR is estimated.

24
App
Load = (PRate × App(hour)), App(hour) = [1/0] (2)
hour=1
P AR = (M ax(LoadApp )/Avg(LoadApp )) (3)

Table 1. Appliances classification
Appliances class Appliances Power rating Operational

(kWh) time (hours)
Shiftable Vacuum cleaner 1.2 6
Electric water heater 2.6 8
Water pump 1 8
Dish washer 2.5 4
Steam iron 1.2 3
Refrigerator 0.225 20
Air conditioner 1.5 14
Non-Shiftable Washing machine 3 5
Tumble dryer 3.3 4
Fixed Oven 1.23 4
Blender 0.3 2
Ceiling fan 0.1 12
Desktop PC 0.3 10
TV 0.3 9
Laptop 0.1 8
4 Scheduling Techniques
Many of the mathematical techniques are presented for scheduling of residen-
tial appliances; MILP, DP, MINLP etc. However, due to slow computation for
large number of appliances consumer demand is not satisfied. Therefore, meta-
heuristic algorithms are used for efficient optimization of home appliances to
reduce the electricity price and manage the load between peak and non-peak
hour. In this paper, we choose some meta- heuristic algorithm SBA and FA to
propose our algorithm RUOA for optimization of energy consumption pattern
of appliances. Algorithms are described further in below subsection.
4.1 Strawberry Algorithm

SBA is a nature based meta-heuristic algorithm of strawberry plant presented in
[19]. These plants grow through runner because of their intelligent nature; they
get their food from nutrient, light and water. If a plant found a good location
for enough food, then it will never move from this location. In another case, if
plant is placed at a location where it found not enough food for its growth then
it try to finding some optimal location by sending long runner for its survival.
The plant propagates some long runner to calculate maximum optimal solution
for survival in Eq. (4). It is hard to provide enough sources to runner especially
when plant is located at place where no good resources for survival. Natural
resources decide the location for plant that its good or not.
r2 = [r1 + drunner ∗ (rand(m, N ) − 0.5)r1 + droot ∗ (rand(m, N ) − 0.5)]. (4)

r1 = ul + (uh − ul) ∗ (rand(m, N ). (5)

SA generates population in form of 1 and 0 for optimization of appliances
using Eq. (5). Every binary number represents a solution for respective time
slot and binary representation of number show the ON/OFF status of home
appliances. The SBA searches for local best solution of minimizing cost and
PAR through runner on the basis of fitness function. SBA performs global best
solutions for search in reproduction step from local solutions.
4.2 Firefly Algorithm

FA is developed on the basis of flashing feature of firefly in [20]. As the entire fire-
fly attracts toward the brighter light so brightness is the main objective function
of firefly. The quality of best solution is depend on the intensity of light emitted
by firefly. Every firefly has values for fitness of brightness as solution and attract
toward the brighter firefly. In simple understanding of FA, basic step of rules are
defined as:
• Present brightness objective function.
• Generate firefly population.
• Calculate the light intensity of firefly.
• Calculate attractiveness of firefly.
• Movement toward the brighter firefly.
• Update the light intensity rank and select the best solution.
4.3 Runner Updation Optimization Algorithm

RUOA is our proposed scheme for optimal scheduling of residential appliances. It
is derived from two other above mentioned nature based meta-heuristic scheme
SBA and FA. We built a feature of population updation of runner from SBA in
flashing scenario of FA. In SBA, the plant propagates few runners upon lower
and upper limit of resources to find optimal solution for survival. We get a
refined updated location of resources by comparing rand number of location.
When condition is satisfied then best solution is being updated for population
through runner [19].
While in FA ideal flashing algorithm is perform for the random population
to acquire the possible optimal solution. They get refined value when condition
for brighter firefly is true, which not better for optimal solution. The value of
updated population is unrefined output, because every time FA takes random
number for optimization so in comparison step population may not achieve the
best solution [20]. In RUOA, we refined the updated population with best fitness
feature of runner from SBA to deploy in FA.
5 Simulations Results and Reasoning

In this section, results of proposed algorithm and comparison with other algo-
rithm are evaluated. MATLAB based simulations are analyzed to verify the
Algorithm 1. Algorithm of Runner Updation Optimization Algorithm.

Require: Input: [Initialize random population, MaxIt, Electricity Price;]
1: Determine Length of operational time for appliances and their power rating;
2: for t = 1 : 24 do
3: Evaluate the value of xl and ul limits random population;
4: Evaluate LOT;
5: Find local best solution;
6: for i = 1 : 10 do
7: Determine the Population Size and no of appliances;
8: Identify initial random population between xl and ul;
9: for iter ≤ iterM axIt do
10: Propagate Runner from SBA on search spaces;
11: Identify fitness criteria of notify location;
12: Evaluate the location and update it;
13: end for
14: end for
15: Rank the population;
16: Find the best population for identified location;
17: end for
18: end for
19: Update the LOT of appliances;
20: Decrement from currently selected appliances LOT;
21: end for
22: end for
results of proposed algorithm RUOA. In our work, we compared the result with
existing meta-heuristic SBA and FA algorithm. CPP and RTP pricing scheme
are used for our main objective of cost minimization, PAR reduction and load
management in simulations. Figures 2 and 3 shows the hourly electricity cost
along with plots of pricing signal CPP and RTP.
The hourly load consumptions are shown in Fig. 4 for CPP and Fig. 5 for RTP
along with scheduled and unscheduled pattern of consumption. Our proposed
Fig. 2. Hourly cost with CPP signal

algorithm RUOA show that its peaks for hourly load graph are better than other
meta-heuristic algorithm SBA and FA for both CPP and RTP. We realize that pat-
tern of per hour load consumption of schedule is much better than unscheduled.
Our simulation result depict that proposed algorithm RUOA optimize the hourly
load consumption by shifting load from peak hours to non-peak hours.
Fig. 3. Hourly cost with RTP signall
Fig. 4. CPP hourly load
Fig. 5. RTP hourly load

The result of PAR value generated by SBA, FA and proposed scheme RUOA
for CPP and RTP are shown in Figs. 6 and 7. Both CPP and RTP signals in
figure verify that proposed algorithm outperformed than other algorithm. PAR
is reduced by 70.76% in SBA, 45.21% in FA and 75.1% in proposed RUOA
for CPP, while 65.45% in SBA, 45.2% in FA and 75.02% in proposed RUOA
for RTP with refer to unscheduled PAR. Simulations show that our proposed
scheme RUOA performed in a better way than other algorithm.
The hourly cost of electricity shown in Figs. 8 and 9 for CPP and RTP, which
simulated by SBA, FA and our proposed algorithm RUOA. Simulations depict
the result of our proposed scheme is good enough from other scheme, as schedule
peak of graph show the lower cost at higher price time slot as compared with
unscheduled peaks.
Total electricity cost value is estimated in Figs. 10 and 11 simulated by SBA,
FA and RUOA for CPP and RTP pricing scheme. Simulation shows that perfor-
mance of our proposed scheme is outclassing as compared with SBA and FA. In
Total Cost plots, the total electricity cost is reducing by 24.53% in SBA, 55.4%
in FA and 65.97% in proposed RUOA for CPP, while 19.78% in SBA, 15.21% in
FA and 35.19% in proposed RUOA with respect of unscheduled case. There is
trade-off between cost and waiting time, which mean that our proposed scheme
RUOA is reducing cost while sacrificing consumer waiting time.
Fig. 6. CPP PAR
Fig. 7. RTP PAR

Fig. 8. CPP hourly cost
Fig. 9. RTP hourly cost
Fig. 10. CPP total cost
Consumer waiting time for CPP and RTP pricing signal are presented in
Figs. 12 and 13. In our case, waiting time shows the time limit to turn ON the
home appliances. The consumer comfort level is calculated in term of consumer
waiting time, we have to minimize the waiting time for consumer to maximize the
comfort level and vice versa. The waiting time calculated for CPP is 3.8876 h by
SBA, 6.7775 h by FA and 4.9879 h by proposed RUOA, while for RTP is 3.8409 h
by SBA, 6.7775 h by FA and 4.5403 h by proposed RUOA. Simulation shows that
Fig. 11. RTP total cost
Fig. 12. CPP waiting time
Fig. 13. RTP waiting time
SBA well performed than other algorithm in consumer waiting time. However,
there is trade-off for our proposed RUOA to SBA and FA in waiting time and
total electricity price.
6 Conclusion
An electricity load management is concerned to evaluate at DSM. In this
paper, we develop the scheme for load management from existing meta-heuristic
technique SBA and FA with pricing scheme of CPP and RTP. We proposed an
RUOA algorithm to optimize the residential appliances based on their electric-
ity consumption pattern. We analyzed the performance of our simulation, which
shows that the performance of our proposed technique is better than other meta-
heuristic technique SBA and FA in term of total electricity price and PAR. On
the other hand SBA outperformed for consumer comfort as compared with our
proposed algorithm.
References
1. Gungor, V.C., Sahin, D., Kocak, T., Ergut, S., Buccella, C., Cecati, C., et al.:
Smart grid technologies: communication technologies and standards. IEEE Trans.
Ind. Inform. 7(4), 529–539 (2011)
2. Esther, B.P., Kumar, K.S.: A survey on residential demand side management archi-
tecture, approaches, optimization models and methods. Renew. Sustain. Energy
Rev. 59, 342–351 (2016)
3. Wu, Z., Tazvinga, H., Xia, X.: Demand side management of photovoltaic-battery
hybrid system. Appl. Energy 148, 294–304 (2015)
4. Strbac, G.: Demand side management: benefits and challenges. Energy Policy 36,
4419–4426 (2008). Vancouver
5. Zhu, Z., Jie, T., Lambotharan, S., Chin, W.H., Fan, Z.: An integer linear pro-
gramming based optimization for home demand-side management in smart grid.
In: 2012 IEEE PES Innovative Smart Grid Technologies (ISGT), pp. 1–5. IEEE
(2012)
6. Agnetis, A., de Pascale, G., Detti, P., Vicino, A.: Load scheduling for household
energy consumption optimization. Smart Grid IEEE Trans. 4, 2364–2373 (2013)
7. Colmenar-Santos, A., de Lober, L.N.T., Borge-Diez, D., Castro-Gil, M.: Solutions
to reduce energy consumption in the management of large buildings. Energy Build.
56, 66–77 (2013)
8. Marzband, M., Ghazimirsaeid, S.S., Uppal, H., Fernando, T.: A real-time evalu-
ation of energy management systems for smart hybrid home Microgrids. Electr.
Power Syst. Res. 143, 624–633 (2017)
9. Moon, S., Lee, J.-W.: Multi-residential demand response scheduling with multi-
class appliances in smart grid. IEEE Trans. Smart Grid 9(4), 2518–2528 (2018)
10. Bahrami, S., Wong, V.W.S., Huang, J.: An online learning algorithm for demand
response in smart grid. IEEE Trans. Smart Grid 9(5), 4712–4725 (2018)
11. Bera, S., Misra, S., Chatterjee, D.: C2C: community-based cooperative energy
consumption in smart grid. IEEE Trans. Smart Grid 9(5), 4262–4269 (2018)
12. Behrens, D., Schoormann, T., Knackstedt, R.: Developing an algorithm to consider
mutliple demand response objectives. Eng. Technol. Appl. Sci. Res. 8(1), 2621–2626
(2018)
13. Javaid, N., Ullah, I., Akbar, M., Iqbal, Z., Khan, F.A., Alrajeh, N., Alabed, M.S.:
An intelligent load management system with renewable energy integration for
smart homes. IEEE Access 5, 13587–13600 (2017)
14. Aslam, S., Iqbal, Z., Javaid, N., Khan, Z.A., Aurangzeb, K., Haider, S.I.: Towards
efficient energy management of smart buildings exploiting heuristic optimization
with real time and critical peak pricing schemes. Energies 10(12), 2065 (2017)
15. Khan, I.U., Ma, X., Taylor, C.J., Javaid, N., Gamage, K.: Heuristic algorithm
based dynamic scheduling model of home appliances in smart grid (2018)
16. do Prado, J.C., Qiao, W.: A stochastic decision-making model for an electricity
retailer with intermittent renewable energy and short-term demand response. IEEE
Trans. Smart Grid (2018)
17. Ma, K., Yao, T., Yang, J., Guan, X.: Residential power scheduling for demand
response in smart grid. Int. J. Electr. Power Energy Syst. 78, 320–325 (2016). 129,
452–470
18. Marzband, M., Yousefnejad, E., Sumper, A., Domnguez-Garca, J.L.: Real time
experimental implementation of optimum energy management system in stan-
dalone microgrid by using multi-layer ant colony optimization. Int. J. Electr. Power
Energy Syst. 75, 265–274 (2016)
19. Merrikh-Bayat, F.: A Numerical Optimization Algorithm Inspired by the Straw-
berry Plant. arXiv 2014, arXiv:1407.7399 (2014)
20. Yang, X.-S.: Firefly Algorithms for multimodal optimization. In: International
Symposium on Stochastic Algorithms. Springer, Berlin (2009)
Multiple S-Box Correlation Energy Analysis
Model Based on Particle Swarm Optimization
Wu-jun Yao(&), Hai-bin Yang, Lin Chen, and Bin Wei
School of Cryptographic Engineering, Engineering University of the Chinese

People’s Armed Police, Xi’an 710086, Shaanxi, China
sensornet@163.com
Abstract. Aiming at the problem that the amount of calculation of correlation

energy analysis is too large in the process of attacking multiple S-box corre-
sponding keys, this paper proposes a multiple S-box correlation energy analysis
model based on particle swarm optimization. The particle swarm optimization
algorithm has the characteristics of simple structure, fast search speed and
memory, therefore, our model simultaneously attacks multiple S-boxes, which
can reduce the amount of calculation, thereby achieving the goal of recovering
the key efficiently and correctly. Finally, experimental analysis and verification
results of the DES algorithm indicate that our new energy analysis model has
about 55% improvement in efficiency and 30% improvement in accuracy over
traditional energy analysis models.
1 Introduction
In 2004, Eric Brier proposed a new type of cipher attack method—correlation energy
analysis [1]. This technology uses the Pearson correlation coefficient [2] p in statistics to
analyze the attack, and recover the key to achieve the attack effect. Compared with other
key recovery technologies, it has stronger attack and higher decryption efficiency.
However, in the traditional correlation energy analysis [3], the correlation between the S-
box and the energy curve is poor [4]. Simply attacking single S-box will turn the
remaining S-box into noise, which will affect the efficiency and accuracy of key cracking.
To enhance the correlation between the S-box and the energy curve [5], multiple S-boxes
can be considered as a whole, and be attacked in parallel, but the number of keys guessed
is too large in this situation, which will cause the key to be unrecoverable. For example,
when attacking four S-boxes simultaneously, the number of bits of the key guessed is
four times that of attacking single S-box, who increases exponentially [6].
With the spreading of artificial intelligence optimization algorithm [7], the advan-
tages of some intelligent optimization algorithms are further reflected. The artificial
intelligence algorithm provides convenience of solving the problem of excessive com-
putation in the process of attacking multiple S-boxes simultaneously with its high
computational efficiency and strong reliability. In this paper, a correlation energy anal-
ysis model based on particle swarm optimization (PSO) algorithm is proposed, according
to the characteristic of the encryption key. The DES is used as the experimental object to
carry out the CPA attack experiment. The experimental results show that Our model can
effectively improve the speed of key recovery and has higher accuracy.
https://doi.org/10.1007/978-3-030-15035-8_38
Multiple S-Box Correlation Energy Analysis Model Based on PSO 413
2 Preliminary
2.1 CPA Attack
CPA attack is an extremely effective method of cipher attack [8]. When using this
method to attack a cryptographic device, the attacker does not need to study the device
in detail, and only needs to know the encryption algorithm used. According to the
energy trace generated by the cryptographic device during encryption, the CPA attack
performs key recovery by analyzing the dependence between the energy trace and the
processed data.
2.2 Particle Swarm Optimization

The PSO is an evolutionary algorithm proposed by Kennedy and Eberhart in 1995 [9].
It quickly solves the optimal solution by simulating the foraging of flocks of birds [10].
The algorithm is concise, easy to implement, fast in convergence, less parameter set-
ting, and easy to program [11]. Analogous to the behavior of birds in the search for
food, individuals in the flock follow the following principles: (1) Avoid collisions with
neighboring individuals; (2) Try to keep pace with the individuals around itself; (3) Try
to be close to the group it agrees with.
The PSO algorithm initializes a population during the solution process. The
algorithm first uses the discriminant function to calculate the priority of each individual
position of the population. The closer an individual is to a food, the higher its priority
will be, conversely, the farther away from the food, the higher its priority will be. And
each individual will have its own search area, from which the best solution for each
individual in their respective regions and the global optimal solution over the entire
range can be derived. In the next solution process, iterations are performed again and
again. During the iterative process, the priority of each individual’s location is still
calculated. If the priority of the individual location at this time is higher than the
priority corresponding to the best solution of the current region of the individual, the
optimal solution of the region is replaced, otherwise it remains unchanged. Similarly,
the priority of the individual position is compared with the priority corresponding to the
global optimal solution in the population. If the priority of the individual position is
higher than the priority corresponding to the current global optimal solution, the global
optimal solution is replaced, otherwise it remains unchanged [12]. After each iteration,
the speed and position of each particle is updated according to the formula.
Suppose the search space is D-dimensional and the total number of particles is n.
The ith particle position is expressed as a vector Xi = (xi1, xi2, …, xiD), the past optimal
position in the “flying” history of the i-th particle is Pi = (pi1, pi2, …, piD), where the
past optimal position Pg of the gth particle is the optimal position among all Pi(i = 1, …,
n), the position change rate (speed) of the ith particle is the vector Vi = (vi1, vi2, …, viD).
The position of each particle changes as follows:
xid ðt þ 1Þ ¼ xid ðtÞ þ vid ðt þ 1Þ 1 i n, 1 d D ð1Þ

414 W. Yao et al.
vid ðt þ 1Þ ¼ w vid ðtÞ þ c1 randð Þ ½pid ðtÞ xid ðtÞ þ c2 randð Þ

½pid ðtÞ xid ðtÞ ð2Þ
Where c1 and c2 are positive constants, called acceleration factors; rand( ) is a

random number between [0, 1]; w is called the inertia factor, it is suitable for large-scale
exploration of the solution space when it is large, and it is suitable for small-scale
excavation when it is small. The position range of dth (1 d D) dimension is
[−Xmaxd, Xmaxd], and the speed range is [−Vmaxd, Vmaxd]. If the position and
velocity exceed the boundary in the iteration, the boundary value is taken. The initial
position and velocity of the particle swarm are randomly generated, and then iterated
according to Eqs. (1) and (2) until a satisfactory solution is found. At present, the
commonly used particle swarm algorithm [13] divides the whole particle group
(Global) into several adjacent subgroups with partial particle overlap, and each particle
adjusts the position according to the historical optimal P1 in the subgroup (Local), that
is, Pgd is replaced with Pld in formula (2).
3 Energy Analysis Model Based on PSO
3.1 Advantages of Multiple S-Box Attacks

Our new energy analysis model is based on particle swarm optimization and multi-box
attack. It takes advantages of fast search and efficient of PSO, and converts the
decoding of the recovery key into the position where the individual is to be iterated.
And choose the appropriate fitness function to compare the pros and cons of the
individual’s location. The fitness function F determines how close the guess key is to
the correct key.
For the problem of low efficiency and large computational complexity of traditional
energy analysis models [14], we use multiple S-box correlations and particle swarm
optimization to solve key spaces quickly, and launch attacks on multiple S-boxes
simultaneously. Through experiments, to verify the advantages of our experimental
model in the simultaneous attack of eight S-boxes compared to the traditional attack
model that attacks single S-box. By comparing experiments, it is verified that the
experimental model has the advantage of attacking the eight S-boxes at the same time
compared to the traditional attack model that attacks an S-box.
In this paper, the DES algorithm is used for data encryption, and the correlation
coefficients obtained by attacking a single S-box and multiple S-boxes are simulated
respectively. In a traditional energy analysis model, a single S-box is attacked, and in
this model, eight S-boxes are simultaneously attacked. Figures 1 and 2 show the
relationship between the correlation coefficient and the number of energy curves in the
DES algorithm. The number of curves is the number of energy curves required to
calculate the correlation coefficient. It can be seen that as the energy curve increases,
the correlation coefficient representing the correct key is gradually separated from the
correlation coefficient of the wrong key.
Fig. 1. Energy analysis for single S box
From Fig. 1, we can conclude that there are 64 keys in the attack of a single S-box,
so there are 64 curves. In addition to the separated red curve representing the energy
trace produced by the correct key, the remaining curves are the energy traces produced
by the 63 incorrect keys. Through experiments we can conclude that at the point of
difference of about 550 energy curves, we can separate the correct key correlation
coefficient from the error key correlation coefficient.
Fig. 2. Energy analysis for 8-S boxes
In Fig. 2, the number of energy traces generated is too large due to the attack of 8
S-boxes as a whole. Therefore, the experimental results show only the correlation
coefficients of some energy curves. The red curve separated from the rest of the curve
represents the energy trace produced by the correct key, and the remaining curves are
the energy traces generated by the partial error key. Through the distinguishing points
in the experiment, we can conclude that for an energy analysis attack of 8 S-boxes, only
250 energy curves are needed to distinguish the correct key from the wrong key.
Compared to the traditional energy analysis attack for a single S-box, the energy
analysis model that attacks 8 S-boxes simultaneously reduces the energy curve required
to distinguish the correct key from the wrong key by about 55%.
416 W. Yao et al.
In this paper, we build an attack model by analyzing the Hamming weight of

multiple S-boxes. Through experimental verification, we can conclude that the corre-
lation of the 8 S-box attack models is significantly higher than that of a single S-box.
Our model avoids the energy information of the remaining S-boxes from becoming
noise, and enhances the correlation between energy information and data. It shows that
our scheme significantly improves the SNR [15], making better use of the power
information of each S-box.
3.2 Energy Analysis Attack Based on PSO

It can be seen from the above experiments that the energy analysis of a plurality of S-
boxes has a stronger correlation. However, the increase in correlation is accompanied
by a sharp expansion of the key space. Therefore, we need to solve the problem quickly
by means of the particle swarm optimization algorithm, transform the key space into
the form of coding, and become a specific solution. Each solution is equivalent to an
independent individual in the particle group, so that it can be quickly obtained by
particle swarm optimization.
Our new energy analysis model consists of 5 steps in the build process.
Step 1: Encode the key. The key is encoded in a binary string of 0 and 1, and the
48-bit key of the S box in DES is divided into eight, each having a 6-bit key. Table 1
shows a 48-bit key encoding process, Xi is the key of the ith S-box, Key is the decimal
key of the S-box, and Genome is a binary coded form of the key.
Table 1. Round key encoding

X1 X2 X3 X4 X5 X6 X7 X8
Key 8 16 7 34 33 30 19 1
Genome 001000 010000 000111 100010 100001 011110 010011 000001
Step 2: Population initialization. After encoding the key, we needs to initialize the
particle swarm and determine the size and scale of the initial population. In the process
of initializing the population, if the number of birds at the beginning is too small, the
effect of the final solution may be affected, and a local optimal solution may occur; if
the number of birds is too large, the collection speed and solution speed of the bird
group will be reduced. Therefore, the selection of population parameters needs to be
specifically set for specific problems. Table 2 completes the initialization of m indi-
viduals in a population. Each individual in the initial population is equivalent to a
randomly guessed round key.
Table 2. Population initialization

X1 X2 X3 X4 X5 X6 X7 X8
1 101100 010111 011100 101010 010101 111100 010001 101100
2 011110 010110 101100 010111 000111 001101 101100 110010
… … …. …
m 111010 010111 010001 111110 001010 011001 101001 001111
Step 3: Determine the fitness function and optimize the individual position in the
particle group by calculating the Hamming weight. In each population iteration, you
need to choose a random solution that is closest to the correct key. In this paper, we
chose a discriminant function to calculate the priority of each individual location. The
discriminant function is defined as data input, that is the degree of association between
the Hamming weight and the correct key generated after 8 S-boxes.
F ¼ ðcorrðTrace; sumðHD; 2ÞÞÞ ð3Þ
The higher the degree of association, the closer the current solution is to the actual
solution, and the better the solution. A new energy analysis model can be established
by calculating the Hamming weight of the eight S-boxes. The individual association
value determines the priority of the individual location. For example, Each generation
has 90 individuals who can solve and rank the associated values for each individual
location. Select the individual location with a high association, and use the particle
velocity and neighborhood definitions to update the historical best position (pbest) and
global best position (gbest) operations.
Step 4: Adjust the moving speed and search range of PSO. In each iteration, the
particle velocity and the locality need to be adjusted to better find a reasonable indi-
vidual. The particle’s velocity update is mainly composed of three parts: its own speed
wvk in the previous iteration; the self-cognitive part c1 r1 ðpbestid xk1
id Þ, indicating the
choice of the particle itself, The distance between the current position of the particle
i and it’s optimal position; the social experience part c2 r2 ðpbestid xk1 id Þ, which
indicates the information sharing and cooperation between the particles, the distance
between the current position of the particle i and the best position of the group.
After each iteration speed and phase adjustment of the group, each particle will
generate a motion vector with the inertia vector under the influence of the regional
optimal solution and the global optimal solution, so that the particle moves to the
optimal position and is correctly aligned. The key is further solved.
Step 5: Align the correlation coefficient [16]. After several iterations, compare the
correlation coefficients. If the preset correlation function value is reached, the optimal
position is obtained, and the correct key is restored. Otherwise return to Step 3 and
optimize the individual again.
During the motion, the PSO will ensure that the guess keys in the 8 S-boxes are all
derived from the correct key. Once the correct S-box key is found, the current solution
would be better than that not reaching the correct S-box key. Therefore, during each
iteration, the individual containing the partial correct S-box key position will step
closer to the optimal position. Due to changes in particle velocity and clinical fields, the
location code represents the key can be optimized for a global optimal solution.
Compared with the traditional energy analysis model, our new energy analysis
model is combined with the artificial intelligence algorithm to perform key recovery.
The model will have a qualitative improvement in efficiency and accuracy. In the
process of recovering the key, the new energy analysis model can make better use of
the energy information and enhance the correlation between the energy consumption of
each S-box, so that the energy curve required is greatly reduced.
418 W. Yao et al.
4 Analysis of Experimental Result

4.1 Parameter Verification
Before performing a new energy analysis, we need to adjust the relevant parameters of
the PSO. In the PSO, there are many parameters affecting its performance, in which the
inertia weight w and the acceleration factors c1 and c2 have the greatest influence on it.
Therefore, this paper tests the three parameters w, c1 and c2 in the genetic algorithm.
Fig. 3. Influence of inertia weight and acceleration factor on the number of iterations
From Fig. 3, we can get that under the premise of the initial population fixed, the
inertia weight w and the acceleration factors c1, c2 will affect the number of iterations
required to recover the key [17]. When the inertia weight w = 0.65 and the acceleration
factors c1 and c2 are 1.60, the PSO can achieve better optimization results.
Figure 4 shows the average number of iterations for different population numbers
obtained after 10 calculations when the inertia weight w = 0.65 and the acceleration
factors c1 and c2 are 1.60.
Fig. 4. Number of populations V.S. number of iterations

The trend shown in the figure is generated because as the number of populations
increases, the number of population iterations decreases first and then increases,
eventually returning to stability, thereby responding to changes in information and
population diversity in the particle swarm [18].
Fig. 5. Iteration number V.S. correlation coefficient
The correlation coefficient in Fig. 5 refers to the correlation coefficient of the

optimal individual in the population when reached the corresponding number of iter-
ations. It shows that when our new energy analysis model uses the PSO for energy
analysis, the solution process is the key recovery process. As the number of iterations
increases, the correlation coefficient increases steadily until the maximum correlation
coefficient is reached, and the correct key is recovered.
4.2 Model Evaluation Experiment

After determining the parameters, the inertia weight w = 0.65, the acceleration factors
c1 and c2 are 1.60, and the number of population is 90. The model is evaluated.
Fig. 6. Number of energy curves and correct rate

420 W. Yao et al.
Figure 6 shows the relationship between the energy curve required to recover the
key and the correct rate in the DES encryption algorithm using the correlation energy
analysis method. From the figure we can see that when the energy curve is 600, the
correct rate of the new energy attack model has reached 100%, while the traditional
energy analysis model has a correct rate of about 70%, and the correct rate is 30%.
Through experiments, it is concluded that under the condition of limited energy curve,
the success rate of recovery key recovery of the new energy attack model combined
with PSO has been improved.
5 Conclusion
We proposes a multiple S-box energy analysis attack model based on particle swarm
optimization in this paper. The model can quickly search the key space by using the
particle swarm optimization algorithm, and convert the key recovery into a binary
string solution, which greatly improves the key recovery speed. In the design and
construction process, the model satisfies the requirements of the efficiency and cor-
rectness of the correlation energy analysis attack theory. Compared with the traditional
correlation energy analysis model, the improved energy analysis model makes the
cryptographic device have a better correlation with the actual energy consumption.
Through experimental analysis and verification, we can conclude that the multi-S box
energy analysis model based on particle swarm optimization reduces the demand for
the number of energy curves by about 55% compared with the traditional single S-box
energy analysis model. In the case of a limited number of curves, the key recovery
correct rate is increased by 30%. It explains that our new energy analysis model can
recover the correct key more efficiently and accurately.
References
1. Brier, E., Clavier, C., Olivier, F.: Correlation power analysis with a leakage model. In:
Cryptographic Hardware and Embedded Systems-CHES 2004 (2004)
2. Li, H., He, G., Guo, Q.: Similarity retrieval method of organic mass spectrometry based on
the Pearson correlation coefficient. Chem. Anal. Meterage 24(3), 33–37 (2015)
3. Mizuno, H., Iwai, K., Tanaka, H., Kurokawa, T.: A correlation power analysis counter-
measure for Enocoro-128 v2 using random switching logic. In: 2012 Third International
Conference on Networking and Computing (ICNC) (2012)
4. Zhang, Z., Wu, L., Wang, A., et al.: Improved Leakage Model Based on Genetic Algorithm.
IACR Cryptology EPrint Archive, 2014: 314 (2014)
5. Nakai, T., Shibatani, M., Shiozaki, M., Kubota, T., Fujino, T.: Side-channel attack resistant
AES cryptographic circuits with ROM reducing address-dependent EM leaks. In: 2014 IEEE
International Symposium on Circuits and Systems (ISCAS) (2014)
6. Qiu, W.-X., Xiao, K.-Z., Ni, F., Huang, H.: DES key extension method. Comput. Eng. 37(5),
167–168+171 (2011)
7. Khalid, S., Khalil, T., Nasreen, S.: A survey of feature selection and feature extraction
techniques in machine learning. In: Science and Information (2014)
8. Zhang, H., Zhou, Y., Feng, D.: Theoretical and practical aspects of multiple samples
correlation power analysis. Secur. Commun. Netw. 9(18), 5166–5177 (2016)
9. Kennedy, J., Eberhart, R.: Particle swarm optimization. In: Proceedings of the IEEE
International Conference on Neural Networks, pp. 1942–1948. IEEE (1995)
10. Pathak, V.K., Singh, A.K., Singh, R., Chaudhary, H.: A modified algorithm of particle
swarm optimization for form error evaluation. tm-Tech. Mess. 84(4), 272–292 (2017)
11. Huang, W.-X.: Research on the development of particle swarm optimization. Comput. Eng.
Softw. 35(4), 73–77 (2014)
12. Hemanth, D.J., Umamaheswari, S., Popescu, D.E., Naaji, A.: Application of genetic
algorithm and particle swarm optimization techniques for improved image steganography
systems. Open Phys. 14(1), 452–462 (2016)
13. Kennedy, J., Eberhart, R.C.: A discrete binary version of the particle swarm algorithm. In:
1997 IEEE International Conference on Systems, Man, and Cybernetics. Computational
Cybernetics and Simulation (1997)
14. Hu, M.-Y., Liu, R.-H.: Analysis and research on security of DES algorithm. Acta Sci. Nat.
Univ. NeiMongol 6, 95–99 (2005)
15. Pan, Q., Zhang, L., Dai, G., et al.: Two denoising methods by wavelet transform. IEEE
Trans. Signal Process. 47(12), 3401–3406 (1999)
16. Lopez, M.C., Fabregas, X.: Polarimetric SAR speckle noise model. IEEE Trans. Geosci.
Remote Sens. 41(10), 2232–2242 (2003)
17. Wang, D.-F., Meng, L.: Performance analysis and parameter selection of PSO algorithm.
Acta Autom. Sin. 42(10), 1552–1561 (2016)
18. Duan, X.-D., Gao, H.-X., Zhang, X.-D., Liu, X.-D.: Relations between population structure
and population diversity of particle swarm optimization algorithm. Comput. Sci. 34(11),
164–166 (2007)
Improving Peer Reliability Considering
Jitter Parameter: A Fuzzy-Based System
for JXTA-Overlay P2P System
Yi Liu1(B) , Makoto Ikeda2 , Keita Matsuo2 , Leonard Barolli2 ,

1
ryuui1010@gmail.com
2
Fukuoka Institute of Technology (FIT),
makoto.ikd@acm.org, {kt-matsuo,barolli}@fit.ac.jp
3
Department of Advanced Sciences, Faculty of Science and Engineering,
Hosei University, 3-7-2, Kajino-machi, Koganei-shi, Tokyo 184-8584, Japan
Abstract. In this paper, we propose and evaluate a new fuzzy-based

reliability system for Peer-to-Peer (P2P) Communications in JXTA-
Overlay platform. We consider Jitter (JT) as a new parameter. In our sys-
tem, we considered four input parameters: Data Download Speed (DDS),
Number of Interactions (NI), Sustained Communication Time (SCT) and
Jitter (JT) to decide the Peer Reliability (PR). We evaluate the proposed
system by computer simulations. From the simulations results, we con-
clude that when DDS, NI and SCT are increased, the PR is increased.
But when JTT is increased, the PR is decreased. The simulation results
have shown that the proposed system has a good performance and can
choose reliable peers to connect in JXTA-Overlay platform.
1 Introduction
The Internet is growing every day and the performance of computers is increased
exponentially. However, the Internet architecture is based on Client/Server (C/S)
topology, therefore can not use efficiently the clients features. Also, with appear-
ance of new technologies such as ad-hoc networks, sensor networks, body networks,
home networking, new network devices and applications will appear. Therefore, it
is very important to monitor, control and optimize these network devices via com-
munication channels. However, in large-scale networks such as Internet, it is very
difficult to control the network devices, because of the security problems.
The Peer-to-Peer (P2P) networks will be very important for future dis-
tributed systems and applications. In such systems, the computational burden of
the system can be distributed to peer nodes of the system. Therefore, in decen-
tralized systems users become themselves actors by sharing, contributing and
https://doi.org/10.1007/978-3-030-15035-8_39
Improving Peer Reliability Considering Jitter Parameter 423
controlling the resources of the system. This characteristic makes P2P systems
very interesting for the development of decentralized applications [1,2].
In [1], it is proposed a JXTA-based P2P system. JXTA-Overlay is a mid-
dleware built on top of the JXTA specification, which defines a set of protocols
that standardize how different devices may communicate and collaborate among
them. It abstracts a new layer on the top of JXTA through a set of primitive
operations and services that are commonly used in JXTA-based applications and
provides a set of primitives that can be used by other applications, which will be
built on top of the overlay, with complete independence. JXTA-Overlay provides
a set of basic functionalities, primitives, intended to be as complete as possible
to satisfy the needs of most JXTA-based applications.
In P2P systems, each peer has to obtain information of other peers and
propagate the information to other peers through neighboring peers. Thus, it is
important for each peer to have some number of neighbor peers. Moreover, it is
more significant to discuss if each peer has reliable neighbor peers. In reality, each
peer might be faulty or might send obsolete, even incorrect information to the
other peers. If a peer is faulty, other peers which receive incorrect information on
the faulty peer might reach a wrong decision. Therefore, it is critical to discuss
how a peer can trust each of its neighbor peers [3–5].
The reliability of peers is very important for safe communication in P2P
system. The reliability of a peer can be evaluated based on the reputation and
interactions with other peers to provide services. However, in order to decide the
peer reliability are needed many parameters, which make the problem NP-hard.
Fuzzy Logic (FL) is the logic underlying modes of reasoning which are approx-
imate rather then exact. The importance of FL derives from the fact that most
modes of human reasoning and especially common sense reasoning are approxi-
mate in nature. FL uses linguistic variables to describe the control parameters.
By using relatively simple linguistic expressions it is possible to describe and
grasp very complex problems. A very important property of the linguistic vari-
ables is the capability of describing imprecise parameters.
The concept of a fuzzy set deals with the representation of classes whose bound-
aries are not determined. It uses a characteristic function, taking values usually in
the interval [0, 1]. The fuzzy sets are used for representing linguistic labels. This can
be viewed as expressing an uncertainty about the clear-cut meaning of the label.
But important point is that the valuation set is supposed to be common to the
various linguistic labels that are involved in the given problem.
The fuzzy set theory uses the membership function to encode a preference
among the possible interpretations of the corresponding label. A fuzzy set can be
defined by examplification, ranking elements according to their typicality with
respect to the concept underlying the fuzzy set [6].
In this paper, we propose a fuzzy-based peer reliability system for JXTA-
Overlay P2P platform considering four parameters: Data Download Speed
(DDS), Number of Interactions (NI), Sustained Communication Time (SCT)
and Jitter (JTT) to decide the Peer Reliability (PR).
The structure of this paper is as follows. In Sect. 2, we introduce JXTA-
Overlay. In Sect. 3, we present the proposed fuzzy-based peer reliability system.
424 Y. Liu et al.
In Sect. 4, we discuss the simulation results. Finally, conclusions and future work
are given in Sect. 5.
2 JXTA-Overlay
JXTA-Overlay project is an effort to use JXTA technology for building an over-
lay on top of JXTA offering a set of basic primitives (functionalities) that are
most commonly needed in JXTA-based applications [7–11]. The proposed over-
lay comprises the following primitives:
• peer discovery,
• peer’s resources discovery,
• resource allocation,
• task submission and execution,
• file/data sharing, discovery and transmission,
• instant communication,
• peer group functionalities (groups, rooms etc.),
• monitoring of peers, groups and tasks.
This set of basic functionalities is intended to be as complete as possible
to satisfy the needs of JXTA-based applications. The overlay is built on top of
JXTA layer and provides a set of primitives that can be used by other appli-
cations, which on their hand, will be built on top of the overlay, with complete
independence. The JXTA-Overlay project has been developed using the ver-2.3
JXTA libraries. In fact, the project offers several improvements of the origi-
nal JXTA protocols/services in order to increase the reliability of JXTA-based
distributed applications and to support group management and file sharing.
The architecture of the P2P distributed platform we have developed using
JXTA technology has two main peers: Broker and Client. Altogether these two
peers form a new overlay on top of JXTA. The structure of JXTA-Overlay system
is shown in Fig. 1.
Fig. 1. Structure of JXTA-Overlay system.

3 Proposed Fuzzy-Based Peer Reliability System

To complete a certain task in JXTA-Overlay network, peers often have to interact
with unknown peers. Thus, it is important that peers must select reliable peers
to interact. The number of interactions that a peer has with other peers in
JXTA-Overlay P2P network is a very important factor that affects the peer
reliability. Another important parameter that is connected with peer reliability
is the Sustained Communication time. Selfish peers that benefits from the system
without contributing any resources to the network have a low reliability. Also
data download speed and jitter are important parameters that effect the peer
behavior. Every time a peer joins JXTA-Overlay, parameters are fuzzified using
fuzzy system, and based on the decision of fuzzy system a reliable peer is selected.
After peer selection, the data for this peer are saved in the database as shown in
Fig. 2. The structure of this system called Fuzzy-based Peer Reliability (FPRS)
System is shown in Fig. 3.
Database
Peer Information
Input Parameter
JXTA-Overlay Fuzzy System
Output Parameter
Join Nerwork
Peer
Fig. 2. Interaction of JXTA-Overlay and fuzzy system.
DDS
NI
FPRS PR
SCT
JTT
Fig. 3. FPRS structure.

426 Y. Liu et al.
In this work, we consider four parameters: Data Download Speed (DDS),

Number of Interactions (NI), Sustained Communication Time (SCT) and Jit-
ter (JT) to decide the Peer Reliability (PR). These four parameters are not
correlated with each other, for this reason we use fuzzy system. The member-
ship functions for our system are shown in Fig. 4. In Table 1, we show the fuzzy
rule base of our proposed system, which consists of 81 rules.
The input parameters for FPRS are: DDS, NI, SCT and JT, while the output
linguistic parameter is PR. The term sets of DDS, NI, SCT and JT are defined
respectively as:
DDS = {Slow, M iddle, F ast}

= {Sl, M i, F a};
N I = {Small, M edium, M any}
= {Sm, M e, M a};
SCT = {Short, M edium, Long}
= {Sho, M ed, Lon};
JT = {Low, M iddle, High}
= {Low, M id, Hig}.
and the term set for the output PR is defined as:

⎛ ⎞ ⎛ ⎞
Extremely Bad EB
⎜ Bad ⎟ ⎜ BD ⎟
⎜ ⎟ ⎜ ⎟
⎜ M inimally Good ⎟ ⎜ M G ⎟
⎜ ⎟ ⎜ ⎟
PR = ⎜ ⎟ ⎜ ⎟
⎜ P artially Good ⎟ = ⎜ P G ⎟ .
⎜ Good ⎟ ⎜ ⎟
⎜ ⎟ ⎜ G ⎟
⎝ V ery Good ⎠ ⎝ V G ⎠
V ery V ery Good VVG
Fig. 4. Membership functions.

428 Y. Liu et al.
Table 1. FRB.
No. DDS NI SCT JT PR No. DDS NI SCT JT PR

1 Sl Sm Sho Low BD 41 Mi Me Med Mid PG
2 Sl Sm Sho Mid EB 42 Mi Me Med Hig MG
3 Sl Sm Sho Hig EB 43 Mi Me Lon Low VVG
4 Sl Sm Med Low MG 44 Mi Me Lon Mid VG
5 Sl Sm Med Mid EB 45 Mi Me Lon Hig G
6 Sl Sm Med Hig EB 46 Mi Ma Sho Low VG
7 Sl Sm Lon Low G 47 Mi Ma Sho Mid MG
8 Sl Sm Lon Mid MG 48 Mi Ma Sho Hig BD
9 Sl Sm Lon Hig BD 49 Mi Ma Med Low VVG
10 Sl Me Sho Low BD 50 Mi Ma Med Mid G
11 Sl Me Sho Mid EB 51 Mi Ma Med Hig PG
12 Sl Me Sho Hig EB 52 Mi Ma Lon Low VVG
13 Sl Me Med Low PG 53 Mi Ma Lon Mid VVG
14 Sl Me Med Mid BD 54 Mi Ma Lon Hig VG
15 Sl Me Med Hig BD 55 Fa Sm Sho Low G
16 Sl Me Lon Low VG 56 Fa Sm Sho Mid MG
17 Sl Me Lon Mid PG 57 Fa Sm Sho Hig BD
18 Sl Me Lon Hig MG 58 Fa Sm Med Low VVG
19 Sl Ma Sho Low MG 59 Fa Sm Med Mid G
20 Sl Ma Sho Mid BD 60 Fa Sm Med Hig PG
21 Sl Ma Sho Hig EB 61 Fa Sm Lon Low VVG
22 Sl Ma Med Low VG 62 Fa Sm Lon Mid VG
23 Sl Ma Med Mid MG 63 Fa Sm Lon Hig VG
24 Sl Ma Med Hig BD 64 Fa Me Sho Low VG
25 Sl Ma Lon Low VVG 65 Fa Me Sho Mid PG
26 Sl Ma Lon Mid G 66 Fa Me Sho Hig MG
27 Sl Ma Lon Hig PG 67 Fa Me Med Low VVG
28 Mi Sm Sho Low MG 68 Fa Me Med Mid VG
29 Mi Sm Sho Mid EB 69 Fa Me Med Hig G
30 Mi Sm Sho Hig EB 70 Fa Me Lon Low VVG
31 Mi Sm Med Low G 71 Fa Me Lon Mid VVG
32 Mi Sm Med Mid MG 72 Fa Me Lon Hig VVG
33 Mi Sm Med Hig BD 73 Fa Ma Sho Low VVG
34 Mi Sm Lon Low VVG 74 Fa Ma Sho Mid G
35 Mi Sm Lon Mid G 75 Fa Ma Sho Hig PG
36 Mi Sm Lon Hig PG 76 Fa Ma Med Low VVG
37 Mi Me Sho Low PG 77 Fa Ma Med Mid VVG
38 Mi Me Sho Mid BD 78 Fa Ma Med Hig VG
39 Mi Me Sho Hig BD 79 Fa Ma Lon Low VVG
40 Mi Me Med Low VG 80 Fa Ma Lon Mid VVG
81 Fa Ma Lon Hig VVG
In this section, we present the simulation results for our FPRS system. In our
system, we decided the number of term sets by carrying out many simulations.
Fig. 5. Relation of PR with DDS and NI for different SCT when JT = 10.
430 Y. Liu et al.
From Figs. 5, 6 and 7, we show the relation between DDS, NI, SCT, JT and
PR. In these simulations, we consider the SCT and JT as constant parameters.
In Fig. 5, we consider JT value 10 units. We change the SCT value from 10 to 90
units. When the SCT increases, the PR is increased. Also, when the DDS and
NI are high, the PR is high. In Figs. 6 and 7, we increase JT values to 50 and
90 units, respectively. We see that, when the JT increases, the PR is decreased.
We conclude that the JT parameter has a big effect on the peer selection.
In this paper, we proposed a fuzzy-based system peer reliability. We took into

consideration four parameters: DDS, NI, SCT and JT. We evaluated the per-
formance of proposed system by computer simulations. From the simulations
results, we conclude that when DDS, NI and SCT are increased, the PR is
increased. But, by increasing JT, the PR is decreased.
In the future, we would like to make extensive simulations to evaluate the
proposed system and compare the performance of our proposed system with
other systems.
432 Y. Liu et al.
References
1. Xhafa, F., Fernandez, R., Daradoumis, T., Barolli, L., Caballé, S.: Improvement
of JXTA protocols for supporting reliable distributed applications in P2P systems.
In: International Conference on Network-Based Information Systems (NBiS-2007),
pp. 345–354 (2007)
2. Barolli, L., Xhafa, F., Durresi, A., De Marco, G.: M3PS: a JXTA-based multi-
platform P2P system and its web application tools. Int. J. Web Inf. Syst. 2(3/4),
187–196 (2007)
3. Aikebaier, A., Enokido, T., Takizawa, M.: Reliable message broadcast schemes
in distributed agreement protocols. In: International Conference on Broadband,
Wireless Computing, Communication and Applications (BWCCA-2010), pp. 242–
249 (2010)
4. Spaho, E., Sakamoto, S., Barolli, L., Xhafa, F., Ikeda, M.: Trustworthiness in P2P:
performance behaviour of two fuzzy-based systems for JXTA-Overlay platform.
Soft Comput. 18(9), 1783–1793 (2014)
5. Watanabe, K., Nakajima, Y., Enokido, T., Takizawa, M.: Ranking factors in peer-
to-peer overlay networks. ACM Trans. Auton. Adapt. Syst. 2(3), 1–26 (2007)
6. Terano, T., Asai, K., Sugeno, M.: Fuzzy Systems Theory and Its Applications.
Academic Press Professional Inc., San Diego (1992)
7. Liu, Y., Sakamoto, S., Matsuo, K., Ikeda, M., Barolli, L., Xhafa, F.: Improvement
of JXTA-Overlay P2P platform: evaluation for medical application and reliability.
Int. J. Distrib. Syst. Technol. (IJDST) 6(2), 45–62 (2015)
8. Liu, Y., Barolli, L., Spaho, E., Ikeda, M., Caballe, S., Xhafa, F.: A fuzzy-based
reliability system for P2P communication considering number of interactions, local
score and security parameters. In: International Conference on Network-Based
Information Systems (NBiS-2014), pp. 484–489 (2014)
9. Spaho, E., Matsuo, K., Barolli, L., Xhafa, F., Arnedo-Moreno, J., Kolici, V.: Appli-
cation of JXTA-Overlay platform for secure robot control. J. Mobile Multimed.
6(3), 227–242 (2010)
10. Liu, Y., Sakamoto, S., Matsuo, K., Ikeda, M., Barolli, L.: Improving reliability of
JXTA-Overlay platform: evaluation for e-learning and trustworthiness. J. Mobile
Multimed. 11(2), 34–50 (2015)
11. Liu, Y., Sakamoto, S., Matsuo, K., Ikeda, M., Barolli, L., Xhafa, F.: Improving
reliability of JXTA-Overlay P2P platform: a comparison study for two fuzzy-based
systems. J. High Speed Netw. 21(1), 27–45 (2015)
ANN Based Intrusion Detection Model
Seunghyun Park1 and Hyunhee Park2(B)

1
Graduate School of Information Security, Korea University, Seoul, South Korea
cloakingmode@gmail.com
2
Department of Computer Software, Korean Bible University, Seoul, South Korea
parkhyunhee@gmail.com
Abstract. Anomaly based Intrusion Detection Systems (IDSs) are

known to achieve high accuracy and detection rate. However, a significant
computational overhead is incurred in training and deploying them. In
this paper, we aim to address this issue by proposing a simple Artificial
Neural Network (ANN) based IDS model. The ANN based IDS model
uses the feed forward and the back propagation algorithms along with
various other optimization techniques to minimize the overall compu-
tational overhead, while at the same time maintain a high performance
level. Experimental results on the benchmark CICIDS2017 dataset shows
that the performance (i.e., detection accuracy) of the ANN based IDS
model. Owing to its high performance and low computational overhead,
the ANN with Adam optimizer based IDS model is a suitable candidate
for real time deployment and intrusion detection analysis.
1 Introduction
Network attacks prediction is an important technology of security management.
Existing network intrusion detection and prediction methods did not fully con-
sider the specific environment factors of target network, which may make the
results deviate from the true situation [1]. In addition, the existing intrusion
detection technology uses an intrusion detection technique that detects a traffic
event based on the signature of the attacker and comparing it with the existing
signature. However, as attack techniques evolve and become more sophisticated,
a technique for defending against attacks in real time is becoming necessary. As
diverse attacks occur, the existing pattern-matching based intrusion detection
technology cannot detect attacks with new patterns and has difficulty detect-
ing attacks in real time [2]. Owing to these limitations, the amount of research
on intrusion detection technology for detecting attack patterns based on data
mining is increasing [3].
The performance of intrusion detection systems may be changed through
the selection of a traffic detection model, traffic detection criteria, and detec-
tion factors. In particular, such techniques may use many different resources for
analysis and may result in large numbers of false positives. Therefore, a tech-
nique for reducing the use of resources is required for precise traffic modeling
and analysis that can lower the false-positive rate. In this paper, we propose an
https://doi.org/10.1007/978-3-030-15035-8_40
434 S. Park and H. Park
intrusion detection model based on an artificial neural network (ANN) model. We

model the influencing factors of network attack probabilities on the constructed
ANN model. Experimental analysis shows that our model leads in more accurate
results.
2 Dataset for Network Traffic
In this paper, we use a recent IDS dataset namely CICIDS2017, which covers all
the eleven necessary criteria with common updated attacks such as Denial-of-
service (DoS), Distributed DoS (DDoS), Brute Force, Cross-site Scripting (XSS),
SQL Injection, Infiltration, Port scan and Botnet. The dataset is completely
labelled and more than 78 network traffic features extracted and calculated
for all benign and intrusive flows by using CICFlowMeter software which is
publicly available in Canadian Institute for Cybersecurity website [4]. Unlike
the existing datasets such as DARPA dataset, KDD Cup 99 dataset [5] and
CAIDA dataset, CICIDS2017 dataset provides 14 different attack methods to
which attack detection techniques can be applied. In particular, it includes var-
ious types of DoS attacks that deplete the server resources, such as DoS Hulk,
GoldenEye, Slowloris, and Slowhttptest, as well as a port scan, which is a typ-
ical network attack, and brute force attacks to capture an authority such as
FTP-Patator and SSH-Patator [6].
This study proposes an intrusion detection and prediction model applying the
machine learning algorithm based on 78 data features using the above-mentioned
CICIDS2017 dataset1 [6,7].
3 ANN Based Network Traffic Prediction

We test the 78 extracted features using the Pearson correlation analysis to select
the best short feature set for each attack which can be best detection features
set for each attack. Then, we examine the performance and accuracy of the
selected features with ANN machine learning model. In this paper, we use the
feed forward and the back propagation algorithms to develop an ANN based
IDS model. The hyperparameters used in applying the learning algorithm are
defined in Table 1.
Detection accuracy is used as parameters for evaluation of different intru-
sion detection models. Accuracy is defined as the fraction of elements correctly
classified as true alarms out of all the elements the intrusion detection model
classified as positive elements in the dataset.
TP
Acc = , (1)
TP + FP
where TP and FP represent the True Positive and False Positive, respectively.
1
https://www.unb.ca/cic/datasets/ids-2017.html.
ANN Based Intrusion Detection Model 435
Table 1. Common hyperparameters
Parameters Value
Independent variable y
Dependent variables x, {x = 1, ...,39}
Learning rate 1, 0.1, 0.01, 0.001, 0.0001
Optimizer Adam, Adagrad, RMSProp, Gradient Descent, Momentum
Hidden dimension 256, 512
Dropout 1, 0.9, 0.8, 0.7
In this study, the abnormal traffic prediction model applying the deep learn-
ing algorithm is implemented using the CICIDS2017 dataset provided by the
CIC (Canadian Institute for Cybersecurity) as mentioned above, and the model-
ing results are compared [5]. Through deep-learning based modeling, the normal
traffic and attack traffic of the CICIDS2017 data are predicted, and finally, which
attack will occur in the attack traffic is predicted. The experimental environment
used to evaluate the performance of the model is outlined with the GeForce
GTX1070 8 GB for GPU, Python 3.7, and TensorFlow 1.9. The CICIDS2017
dataset is composed of 2,830,743 records, of which 70% is used for training,
20% for evaluation, and 10% for validation. A total of 39 independent variables
are extracted by conducting feature-selection based on the Pearson correlation
coefficient from the 78 features representing the data characteristics. Selected
39 independent variables are below for input dataset: Bwd Packet Length Std,
Bwd Packet Length Max, Bwd Packet Length Mean, Avg Bwd Segment Size,
Packet Length Std, Max Packet Length, Packet Length Variance, Fwd IAT Std,
Packet Length Mean, Average Packet Size, Idle Max, Idle Mean, Flow IAT Max,
Fwd IAT Max, Idle Min, Flow IAT Std, Min Packet Length, Bwd Packet Length
Min, Fwd IAT Total, Flow Duration, FIN Flag Count, PSH flag Count, Flow
IAT Mean, Bwd IAT Std, Fwd IAT Mean, URG Flag count, Destination Port,
ACK Flag Count, Fwd Packet Length Min, Bwd IAT Max, Idle Std, Init Win
bytes backward, Fwd Packet Length Mean, Avg Fwd Segment Size, SYN Flag
Count, Fwd PSH Flags, Down/Up Ratio, Fwd Packet Length Max, and Fwd
Packets/s.
Furthermore, 8 dependent variables for detecting attacks are extracted,
excluding some attacks that have an extremely small dataset only to a traffic
imbalance. Selected 8 dependent variables are below for output dataset: Benign,
DDoS, DoS GoldenEye, DoS Hulk, DoS Slowhttptest, DoS Slowloris, FTP Pata-
tor, and Portscan.
To validate the feature selection and traffic imbalance, the prediction results
are compared with the results using the complete datasets without an extraction
process. The 39 independent variables selected as input are regularized using
StandardScalar, and one-hot encoding is conducted for the 9 dependent variables
selected as output.
436 S. Park and H. Park
Fig. 1. Simulation result of ANN model with GD and Adam optimizers.
To obtain the optimal hyperparameters for ANN modeling, the model is trained
while varying the learning rate and dropout. As a result, the optimal result is
obtained when the learning rate is 0.001 and the dropout is 1. Figure 1 shows
the results of applying the Gradient Descent (GD) and Adam optimizers to the
above hyperparameters. According to the modeling results of an ANN using
the GD optimizer, the prediction accuracy is 98% or higher for most attacks.
However, for FTP-Patator attacks, the sensitivity is 0, which indicates that
the ANN modeling cannot predict certain attacks that show a class imbalance.
Consequently, the results of applying other optimizers are analyzed in addition
to ANN modeling. Although only the optimizer is changed, the class imbalance
problem in certain attacks (e.g., FTP-Patator) disappeared. When the Adam
optimizer is applied to ANN modeling, the prediction ratios are constant for all
attacks in general.
5 Conclusion
In this paper, we proposed an anomaly detection model based IDS using the
ANN learning algorithm. Experimental results on the benchmark CICIDS2017
dataset shows that its performance is comparable to that of the ANN with GD
optimizer and ANN with Adam optimizer. However, since the ANN with Adam
optimizer based IDS model uses only a single hidden layer, its computational
overhead is comparatively less than that of the ANN with GD optimizer based
IDS models. Therefore, the ANN with Adam optimizer based IDS model is an
ANN Based Intrusion Detection Model 437
appropriate candidate for real time deployment and intrusion detection analysis.
For our future work, we aim to fine tune various parameters of the proposed IDS
model to further enhance its performance.
Acknowledgements. This work was supported by the National Research Founda-

tion of Korea (NRF) grant funded by the Korea government (MSIT) (No. NRF-
2017R1C1B5017556).
References
1. Roesch, M.: Snort - lightweight intrusion detection for networks. In: Proceedings
13th USENIX Conference on System Administration, pp. 229–238 (1999)
2. Mukkamala, S., Janoski, G., Sung, A.: Intrusion detection using neural networks
and support vector machines. In: Proceedings International Joint Conference on
Neural Networks, pp. 1702–1707 (2002)
3. Yu, Y., Wu, H.: Anomaly intrusion detection based upon data mining techniques
and fuzzy logic. In: Proceedings IEEE International Conference on Systems, Man,
and Cybernetics, pp. 514–517 (2012)
4. CICFlowMeter for network traffic generator and analyser. Canadian institute for
cybersecurity (CIC) (2017). https://www.unb.ca/cic/research/applications.html
5. Tavallaee, M., Bagheri, E., Lu, W., Ghorbani, A.: A detailed analysis of the KDD
CUP 99 data set. In: Proceedings IEEE Symposium on Computational Intelligence
for Security and Defense Applications, pp. 1–6 (2009)
6. Sharafaldin, I., Lashkari, A., Ghorbani, A.: Toward generating a new intrusion detec-
tion dataset and intrusion traffic characterization. In: Proceedings 4th International
Conference on Information Systems Security and Privacy, pp. 108–116 (2018)
7. Shiravi, A., Shiravi, H., Tavallaee, M., Ghorbani, A.: Toward developing a systematic
approach to generate benchmark datasets for intrusion detection. Comput. Secur.
31(3), 357–374 (2012)
Telecommunication Networking,
Implementation of a Transnational Testbed
and Web UI System with Layer3 SDX
Wun-Yuan Huang1(&), Hui-Lan Lee1, Ta-Yuan Chou1,

Te-Lung Liu1, Fei Yeh2, Jim Hao Chen2, and Joe Mambretti2
1
National Center for High-Performance Computing, NARLabs, Taipei, Taiwan
{wunyuan,gracelee,1203053,tlliu}@narlabs.org.tw
2
International Center for Advanced Internet Research, Northwestern University,
Evanston, USA
{fyeh,jim-chen,j-mambretti}@northwestern.edu
Abstract. An Internet eXchange Point (IXP) is a central aggregation point

where Internet Service Providers interconnect with each other using the Border
Gateway Protocol (BGP) to peer with Autonomous System (AS). An IXP is also
the most critical part of the Internet’s infrastructure. By incorporating Software
Defined Network (SDN) technology, Software Defined eXchanges (SDXs) can
provide a much more intelligent mechanism in network management and more
efficient, flexible and dynamic network resource provisioning. In this paper, we
describe an adaptive network transmission testbed based on SDN-IP that pro-
vides efficiency, elasticity, security, and convenience. SDN-IP is an application
built on an open-source ONOS, and it plays an important role for intercon-
nection among heterogeneous networks. Using this approach, we exploit the L3
routing mechanism of SDN-IP for inter-domain communication, interconnecting
several research sites. Besides the testbed deployment, we implemented a
Web UI system for this testbed to decrease management overhead.
1 Introduction
Today, many applications and services are developed on the Internet, including those
supported by cloud computing and big data. They are the most popular technologies, so
many people use or study them. To satisfy the increasing number of service requests,
service providers have to extend their infrastructure and build many datacenters across
the world to enhance service capability and redundancy. However, when creating new
datacenters, they encounter problems, such as lack of IP address, management oper-
ations among datacenters, bandwidth assignment, quickly shifting traffic patterns, and
security. To respond to these issues, a novel network architecture may be useful.
Software Defined Network (SDN) is a popular current research topic. Compared
with traditional architecture, SDN is differentiated because the control plane and the
data plane are separated. The architecture of SDN is shown in Fig. 1(a). To implement
the control plane, an SDN controller is installed on a server, for example, ONOS (Open
Network Operating System) [1], OpenDayLight (ODL) [2], and Ryu [3]. The SDN
switches support the data planes. When the first packet enters an SDN switch, it will be
forwarded to the SDN controller. The SDN controller has rules that decide policies and

https://doi.org/10.1007/978-3-030-15035-8_41
442 W.-Y. Huang et al.
writes flow entries to SDN switches. Consequently, the subsequent packets can follow
the appropriate policy to arrive at the right destination. Furthermore, the SDN network
provides for network programmability. SDN can rapidly satisfy the requests of network
service deployments and provide dynamic updates about network status, automatic
network provisioning, and network virtualization functions [4–6].
An Internet eXchange Point (IXP) is able to provide a connection between the
Internet Service Provider (ISP) and the Content Delivery Network provider (CDN). It
can ensure connection quality and decrease network connection costs and end-to-end
latency. In [7, 8], the authors apply SDN to IPX architecture, called an SDN eXchange
point (SDX).
In [7], authors propose an SDX architecture which consists of ONOS SDN and an
IPX. In this SDX, an ONOS application, SDN-IP [9], can interconnect IP and SDN
domains seamlessly. While the SDN-IP receives Border Gateway Protocol (BGP) in-
formation from IP domains, it will update the routing information base (RIB) within
ONOS. Then, ONOS is triggered to write transfer rules within SDN switches.
In this paper, we will describe how to deploy SDN-IP in an academic research
network. Furthermore, we will describe an implementation of a Web UI system for
SDN-IP, which can help manage the SDX network.
The remainder of this paper is organized as follows. Section 2 introduces the
research background, including ONOS, SDN-IP, and Quagga [10]. The deployment of
SDX architecture and implementation of the Web UI system for SDN-IP are described
in Sect. 3. Experiments and results are demonstrated in Sect. 4. The conclusions are
given in the last section.
Fig. 1. (a) SDN architecture. (b) Quagga architecture.
2 Background
In our proposed SDX and system, three applications, such as ONOS, SDN-IP, and
Quagga, are the most important. To explain more clearly, the implementation of the
proposed SDX is introduced here.
Implementation of a Transnational Testbed 443
2.1 ONOS
ONOS is an open-source operating system for an SDN control plane, which supports
the current infrastructure of datacenter virtualization and cloud computing. Moreover, it
is production SDN control plane operating system. ONOS has many applications,
excellent modularization, and support both southbound and northbound APIs. It has
higher scalability, availability, and performance than earlier SDN controllers, such as
NOX [11], POX [12], Beacon [13], and Floodlight [14].
ONOS is distributed as a cluster architecture. If a failure happens to one ONOS
implementation, it will follow pre-defined standard procedures to transfer all data to
other implementations. The Network data in SDN is stored on more than one ONOS
system. Each ONOS implementation has specific tasks to process. Thus, the fault
tolerance mechanism of control plane and network management in the large-scale SDN
network can be enhanced.
The ONOS northbound interface supports the Rest API [15] and Intent [16]
framework. Users can communicate with ONOS by them. For example, if users
implement a REST (Representational State Transfer) API related to devices manage-
ment or an Intent app about routing, they can ask ONOS to manage devices with the
REST API, or add/remove flow entries within SDN switches via the Intent app. The
southbound interface supports P4 [17], OpenFlow, NETCONF [18], TL1 [19], SNMP
[20], and RESTCONF [21] protocols so that ONOS can manage devices that support
those protocols, including, white-box switches, and traditional network devices.
2.2 Quagga
Quagga is an open-source routing software suite under terms of the GNU General
Public License (GPL). It can be installed on Linux as a software-based router. The
architecture of Quagga is shown in Fig. 1(b). It is divided into three parts:
• Routing programs (red blocks): supporting routing protocols, including OSPFv2,
OSPFv4, RIPv1, RIPv2, ISIS, and BGPv4.
• Zebra core program (blue blocks): an abstraction layer between the Linux core and
the router.
• Vtysh (virtual teletype shell) (green blocks): the Telnet server is implemented in
routing programs. When a routing program starts, it will specify a port number to
listen, such as 2605. Therefore, users can login to the server to set the routing
configuration via this port.
The core daemon of Zebra is Zserv, which runs on the server side of Unix or TCP
sockets. It can communicate with a client (denoted as zclient API) on the router to
receive events, such as add, modify, and delete. Zebra runs on the user space of Linux.
By Netlink communication mechanism, it updates routing information recorded within
the routing program to the routing table on kernel space of Linux to ensure all routing
entries in the routing table is up-to-date.
3 Implementation
For this research project, we built an SDN-IP platform. It is a layer3 SDX prototype. In
this section, we illustrate the architecture and how it interconnects with multiple
international research networks. In addition, we describe the implementation details for
the SDN-IP Web UI system.
3.1 SDX Network Architecture and Deploy Configuration

The SDX architecture used for the SDN-Testbed is shown in Fig. 2. Components
deployed in the SDX environment include ONOS controller, Quagga, and SDN
switches. An SDX controller can be a cluster formed with one or more ONOS instances
and then SDN-IP can be run on it. Quagga, a software router, is a BGP speaker
responsible for establishing iBGP routes with SDN-IP within SDX. It can also peer
with traditional IP network domain routers to establish eBGP routing path among
different ASs. All peering information among ASs is exchanged through the SDN
switches.
To ensure extensibility of the SDN control plane, we deploy a management switch
(Mgnt SW) to exchange iBGP information between SDN-IP and Quagga. After
Quagga receives eBGP from other AS, it creates iBGPs based on routing information,
and sends it to SDN-IP. Then these iBGPs are converted to intent to ask ONOS to
install flow entries into the SDN switch data plane.
In the IP network domain, we designed a Local Router to connect with Quagga of
SDX. There are two basic ways for clients to communicate with each other. SDX and
non-SDX. The former means the clients at two ends can communicate directly by
connecting with the SDX through Local Router. The latter means they can connect to
the Internet with Border Router within a LAN.
For the ONOS deployment in this SDX architecture, it is necessary to activate
SDN-IP and deactivate all forwarding apps, such as org.onosproject.fwd. Furthermore,
a network-cfg.json file is required. Figure 3 is a network-cfg.json file written in JSON
format, which contains ports, devices, and BGP section.
• ports section: configurations of the ports connected with ASs have to configure IP,
MAC, and name for peering with ASs.
• devices section: configurations of ONOS to connect to the device.
• BGP section: configurations of the name, conntectPoint, and peers of BGP speaker.
We can reboot ONOS or input “onos-netcfg localhost network-cfg.json” to load
this file. Then SDN-IP can create Single-Point to Single-Point intents.
Fig. 2. Architecture for SDX and IP network.
Fig. 3. Network-cfg.json file contents.
3.2 SDX International Connection

We built an SDX based on the SDN-IP Testbed in the National Center for High-
Performance Computing (NCHC) and connected it with National Chiao Tung
University (NCTU) and National Cheng Kung University (NCKU) in Taiwan. Besides
these two universities, we also looked for an opportunity for extending the testbed
internationally.
Since the next generation network environment will be composed of many different
network technologies, we provide options for heterogeneous network connections,
which enable the network environments of different organizations to use this experi-
mental environment. Figure 4 shows the international connection in the SDX archi-
tecture. NCHC Hsinchu branch, NCHC Tainan branch, and NCKU are interconnected
by Layer2 VPN on Taiwan Advanced Research and Education Network (TWAREN)
VPLS backbone. NCTU connects with both NCHC Hsinchu branch and NCHC Tainan
branch by TWAREN SDN Layer2 VPN service. ASTI, HUST, and OOL connects to
NCHC Hsinchu branch with the VPN Tunnel.
Currently, TWAREN is an official member of the GLIF (Global Lambda Interated
Facility) Open Lambda Exchange (GOLE) consortium. TWAREN connects to Star-
Light International/National Communications Exchange Facility in Chicago and to
PacificWave in California with two International LightPaths. GOLEs, TaiwanLight,
StarLight, and PacificWave are all Internet exchanges. The International Center for
Advanced Internet Research (iCAIR) at Northwestern University is connected to
NCHC via StarLight. NICT is connected to PacificWave through TransPac and to

PacificWave, which also connects to NCHC. KREONet connects to NCHC through
KRLight and StarLight. RNP connects to AmPath and also to PacificWave. Finally,
PacificWave supports a connection to the NCHC SDN-IP Testbed.
Fig. 4. International connection in the SDX architecture.
3.3 Implementation of Web UI System

The Web UI system is divided into three parts: the ONOS northbound interface, Web
Server, and Web UI. The System diagram is shown in Fig. 5(a). In the ONOS
northbound interface, we implement the REST API to enable Web Server to com-
municate with ONOS to obtain data about SDN-IP. The operation procedures of this
system are as follows:
• Step 1. After the user launches the browser and inputs the server URL, the browser
sends the request to the Web Server.
• Step 2. When the Web Server receives the request, it asks ONOS to obtain the
requested data.
• Step 3. When the REST API receives a request from a Web Server, it asks ONOS
for the information it needs.
• Step 4. ONOS sends data back to SDN-IP.
• Step 5. The SDN-IP translates the data to the JSON-formatted string and then sends
it back to Web Server.
• Step 6. The Web Server deals with the data and then transmits it to Web UI.
The Web Server is responsible for retrieving data with ONOS and then sending it
back to the Web UI in JSON format. The data obtained from ONOS are used for two
purposes: SDN-IP information display and SDX network topology. For SDN-IP
information display, it has three types of data as follows:
• BGP-Neighbors: show the BGP information of the local site (SDN-IP) and remote
site (Quagga connected with SDN-IP), such as BGP version, BGP ID, and AS
number.
• BGP-Routes: show all information of the routing table.
• BGP-Speakers: show information of all SDN-IP Speakers, such as speaker name,
VLAN id, device id of attachment point, and peers.
After the Web Server obtains the three types of data, it translates these data into a
JSON-formatted string and transfer to the Web UI system. Next, the Web UI fetches
information from the JSON-formatted string with a JSON Parser coded in JavaScript
and displays it in a tabular format.
For the SDX network topology display, the Web Server fetches the AS path from
BGP-Routes and then fetches data of nodes and links. After the Web Server obtains all
data of nodes and links, it sends these data to the Web UI in a JSON-formatted string so
that the Web UI can display a global topology graph. Figure 5(b) shows the pseu-
docode of the algorithm of fetching these data. We present an example to illustrate this
algorithm. There are 2 SDN-IP sites, and they are A (ASN65000) and B (ASN65111)
respectively. The steps of the algorithm are as follows:
• Step 1. Obtain BGP-Routers data from A site, and obtain the AS path. This AS path
has 2 items and they are [65000] and [65000, 65111] respectively.
• Step 2. We obtain the item [65000] and obtain the ASN65000 from this item.
• Step 3. Since no ASN65000 exists in the nodesArray, it will be added to the
nodesArray.
• Step 4. Obtain the next ASN (denotes as nextASN).
• Step 5. Since nextASN is empty, no link is established.
• Step 6. Repeat step2 to step5. A new node (ASN65111) is added in nodesArray and
a link (ASN65000-ASN65111) is also added in linksArray.
• Step 7. Repeat step 1 to step 6 to fetch nodes and links from the ASN path of B site
until all SDN-IP sites are traversed.
Fig. 5. (a) Execution procedure of web UI system. (b) Pseudocode of algorithm for fetching
nodes and links.
4 Testbed in Web UI
After we completed the deployment of SDN-IP among academic research organization

in both domestic and international domains, the SDX could exchange routing and
forward packets. With this SDX, IP domains can reduce connection delays across
different countries and enable an SDN intelligent feature to provide a flexible appli-
cation service.
Now, we can utilize the proposed Web UI system to see information about SDN-IP
and topology in this SDX. For the sake of simplicity, we only expose some nodes in the
Web UI system. First, we deploy the Web Server software for this system to serve HTTP
requests, such as Apache Tomcat. Then, users can launch a browser and input URL to
browse the portal page of this system. Figure 6(a) is the snapshot of the portal page. It
has options to enable users to input one or multiple SDN controller IP. Here, we input
the SDN controller IP of NCHC Hsinchu and Tainan branch to obtain information about
SDN-IP within these SDN domains. Figure 6(b) demonstrates the information page for
both two domains. Users can click the site tabs in the upper left corner to display the
SDN-IP information for this site and see BGP-Routes, BGP-Neighbors, and BGP-
Speakers information from three tables. If a cell has more than one data in a table, this
field will show the amount of data and add a hyperlink behind this number to expose
these data in JSON-formatted strings, such as PathSegment and Peers cell.
Next, users can click the menu icon on the top of the right-hand side corner to
expose a hyperlink of topology. When we click this hyperlink, it will redirect users to
view the page of topology display. The topology is shown in Fig. 6(c).
As shown in Fig. 6(c), NCHC-TC is connected with NCKU, NICT attached with
NCHC-HC, NCTU it connected with NCHC-TN and NCHC-HC in the topology
window. It is similar to the SDX we built. In the topology window, all nodes and links
are drawn based on each ASN path obtained from BGP-Routes information of SDN-IP.
If we obtain information from one SDN, this node will be marked in a yellow color in
this topology. Others are marked in blue color.
In addition to the topology display, we also implemented a function to expose the
routing path, as shown in Fig. 7. When one clicks the NCHC-HC node, a routing path
window will be exposed. If we move the mouse over a path entry, a routing path will be
exposed in the topology. With this function, users can easily know the routing path
between both two sites. Consequently, it can enhance the routing management in this
SDX.
Fig. 6. (a) Portal page. (b) Information page. (c) Topology page.
Fig. 7. Routing path display.
In this paper, we describe an SDX network architecture and visualized platform based
on SDN-IP technology and Web UI system. We deployed an SDX on TWAREN for
international nodes. In the SDX deployment, we deployed ONOS and launched SDN-
IP on it to interconnect IP domains connected with the SDX.
We successfully demonstrated this SDX in a meeting of a research consortium -
Collaborations to Enable Transnational Cyberinfrastructure Applications (CENTRA).
Next, we will use intelligence and automation technology of SDN to build the first
operational level SDX in the country to provide a network architecture for intercon-
nected among heterogeneous domains. In the future, we will deploy Scalable Amplified
Group Environment 2 (SAGE2) [22] on the multinational transnational IP network
domain as the network environment for disaster prevention telemetry and smart city
application. Moreover, we will add functions in the Web UI system for network control
and management so that SDX can be more scalable and flexible.
References
1. ONOS. http://onosproject.org/
2. OpenDaylight. https://www.opendaylight.org/
3. Ryu. http://osrg.github.com/ryu/
4. McKeown, N., Anderson, T., Balakrishnan, H., Parulkar, G., Peterson, L., Rexford, J., et al.:
OpenFlow: enabling innovation in campus networks. ACM SIGCOMM Comput. Commun.
Rev. 38(2), 69–74 (2008)
5. ON Foundation: “Software-defined networking: The new norm for networks,” Open
Networking Foundation, Tech. Rep. (2012)
6. Astuto, B.N., Mendonca, M., Nguyen, X.N., Obraczka, K., Turletti, T.: A survey of
software-defined networking: past, present, and future of programmable networks. IEEE
Commun. Surv. Tutorials 16(3), 1617–1634 (2014)
7. Gupta, A., Vanbever, L., Shahbaz, M., Donovan, S., Schlinker, B., Feamster, N., Rexford, J.,
Shenker, S., Clark, R., KatzBassett, E.: SDX: a software defined internet exchange.
ACM SIGCOMM (2014)
8. Lin, P., Hart, J., Krishnaswamy, U., Murakami, T., Kobayashi, M., Al-Shabibi, A., Wang, K.C.:
Seamless interworking of SDN and IP. ACM SIGCOMM Comput. Commun. Rev. 43(4), 475–
476
9. SDN-IP. https://wiki.onosproject.org/display/ONOS/SDN-IP
10. Quagga Routing Suite. http://www.nongnu.org/quagga/
11. Gude, N., Koponen, T., Pettit, J., Pfaff, B., Casado, M., McKeown, N., Shenker, S.: NOX:
towards an operating system for networks. ACM SIGCOMM Comput. Commun. Rev. 38
(3), 105–110 (2008)
12. McCauley, M.: POX (2009). https://github.com/noxrepo/pox
13. Erickson, D.: The beacon OpenFlow controller. In: Proceedings of the Second ACM
SIGCOMM Workshop on Hot Topics in Software Defined Networking, ser. HotSDN 2013,
pp. 13–18. ACM, New York (2013)
14. Floodlight, an open SDN controller. http://www.projectfloodlight.org/floodlight/
15. REST API. http://www.restapitutorial.com/
16. Intent Framework. https://wiki.onosproject.org/display/ONOS/Intent+Framework
17. P4 Language Consortium. https://p4.org/
18. Network Configuration Protocol (NETCONF). https://tools.ietf.org/html/rfc6241
19. Transaction Language 1 (TL1) wiki. https://en.wikipedia.org/wiki/Transaction_Language_1
20. Simple Network Management Protocol (SNMP). https://www.ietf.org/rfc/rfc1157.txt
21. RESTCONF Protocol. https://tools.ietf.org/html/rfc8040
22. SAGE2. http://sage2.sagecommons.org
The Case Study of Software Build-in Design
Based on Quality Factors and FMEA
Meng-Ling Hsieh(&), Wei-Tsen Lin, Suhan Yu, Yi-Chi Chen,

Jung-Shan Lin, and Lin-Hui Nung
Chunghua Telecom Laboratories, No. 99, Dianyan Road, Yangmei District,

Taoyuan City 326, Taiwan (R.O.C.)
mlhsieh@cht.com.tw
Abstract. In order to improve the quality and stability of software, software

quality factors must be considered in the early stages of system development as
in [1]. Moreover, in the design stage, we use Failure modes and effects analysis
(FMEA) to analyze the abnormal situation. Different to traditional exception
handling, FMEA is a systematic way of identifying failure modes of a system,
and take actions to eliminate or reduce failures before the failures occur as in [2].
Consequently, the cost and rework of software development could be reduced,
and the defect removal efficiency (DRE) could be increased. Based on McCall’s
software quality factors and FMEA, we derive the quality problems that may be
encountered in the software development stage and the failure scenarios that
may be confronted during the operation phase. Furthermore, we analyze the
failure mode and calculate their risk priority number (RPN). Following the
systematic engineering thinking as in [3], we conclude the early design issues,
and propose the built-in features which are worth to be designed in the early
stage. Finally, we illustrate the multiple effects after its implementation.
1 Introduction
As the software requirements change more and more frequently, the quality standards
and methods for improving software are also evolving as in [4].
The design of considering the issues that might occur in the development/test/
operation stage and incorporating related functions into software architecture in the
early development stage as in [1, 3] is called Built-in Design. The Built-in Design
contains functions such as flexible functional expansion capability, scalable operational
mechanism, and software maintainability, etc. and designer of software should choose
among them with consideration of importance and feasibility to the software and the
schedule of the project.
2 Built-in Design Considerations
2.1 Software Quality Factors

Based on McCall’s software quality factors as in [5], there are three quality measure
index which are Product Operation, Product Revision, and Product Transition that
contain eleven quality factors in total as shown in Fig. 1.

https://doi.org/10.1007/978-3-030-15035-8_42
452 M.-L. Hsieh et al.
Fig. 1. McCall’s software quality factors
Product Operation. As shown in Table 1, Product Operation contains five factors

which are Correctness, Reliability, Efficiency, Integrity, and Usability. Correctness is
one of the most important quality factors that guarantee that software satisfies user
requirements and has correct output. Reliability guarantees that software produces
output within specified time limit without interruption. Efficiency guarantees higher
throughput under the same system environment.
Product Revision. As shown in Table 1, Product Revision contains three factors
which are Maintainability, Flexibility, and Testability. When software failure occurs,
Maintainability means to identify root cause, find solution, and then remove defects as
fast as possible to reduce the influence to related operations. Flexibility guarantees a
flexible architecture to modify software that is online and operating to meet require-
ment changes. Testability guarantees the correctness, reliability, and efficiency meets
requirement.
Product Transition. As shown in Table 1, Product Transition contains three factors
which are Portability, Reusability, and Interoperability. Portability is important when
we upgrade or migrate hardware system. With higher portability, the resources required
to modify and deploy software can be reduced dramatically. One of the methods to
improve portability is to modularize and abstract software. Reusability guarantees the
reuse of software components that enables us to select needed components and
assembles them flexibly with minimum effort for customization. To improve
reusability, software modularization and de-coupling is required. Interoperability
guarantees lower cost to connect to other software. As the maturity of Internet of
Things, IoT, Interoperability is more and more important. To improve interoperability,
specific protocol or standard or flexible Application Programming Interface, API are
required.
The Case Study of Software Build-in Design Based on Quality Factors and FMEA 453
Table 1. Mccall’s software quality factors

Quality Quality factor Description
measure
index
Product Correctness Ability of software that satisfies the design specification
operation and user requirement
Reliability Ability of software that outputs the result exactly as
expected
Efficiency Resources required performing software execution
Integrity Ability of software without unauthorized access or
modification, including its data and output result
Usability Ability of ease of use, including learning, operating,
inputting data and verifying result
Product Maintainability Resources required for removing defects
revision Flexibility Resources required for modifying software that is online
and operating
Testability Resources required for verifying results exactly as
expected
Product Portability Resources required when upgrade or migrate hardware
transition system
Reusability Ability of software to be used in other software
Interoperability Resources required for connecting with other software
2.2 Failure Mode and Effect Analysis, FMEA

Based on TM Forum’s report, one of the challenge that Telecommunication industry
faces when introducing Artificial Intelligence, AI is out-of-control as in [6].
We can use Failure modes and effects analysis (FMEA) as in [2] to analyze
potential error and abnormal and failure scenarios of software. We can calculate risk
priority number (RPN) by Severity, Occurrence, and Detection and can design mea-
sures in advance to avoid defects or reduce the impact of them as in [5]. Based on
FMEA, this study uses Development Stage Failure Mode and Operation Stage Failure
Mode.
Development Stage Failure Mode. Development Stage Failure occurs when software
development environment changes due to software upgrade or parameter modification
without notifying developers. In some cases, Development Stage Failure occurs when
developers modify software to meet new requirements but fail to maintain existing
functionality.
Operation Stage Failure Mode. Operation Stage Failure occurs as operation executed
with wrong order, operation executed at wrong time, system overloaded as multiple
operation executed simultaneously and operation executed without access. Sometimes,
Operation Stage Failure occurs because of quality failure, such as unexpected result,
interrupted service, low efficiency, and information security issues.
2.3 Early Design

The variable cost will gradually increase as the project progresses as in [7], especially
the cost for issues related to software architecture.
FMEA can be used in every phase of software life cycle, including requirement
analysis, design, development, operation, and maintenance phase. But the design phase
is the most appropriate phase that has the most significant effect to apply FMEA.
Because in the design phase, structural defects can be identified and therefore modi-
fication of design and architecture in the later phase with much higher cost can be
avoided as in [8].
3 Built-in Design Functionality
Based on McCall’s software quality factors and FMEA, we conclude the early design
issues, and propose the Built-in Design which is worth to be designed in the early
stage. We design Built-in Design in three aspects, including Operation Control, Quality
Control, and Software Functionality as shown in Table 2.
Table 2. Built-in design functionality

Aspect Functionality
Operation Control Aspect Operation Control Tool
Quality Control Aspect Environmental Diagnostic Tool
Development Procedure & Checklist
Software Functionality Aspect Regression Test Tool
Components and Modules
1. Data Group Control Module
2. Batch and Parallel Processing Module
3. Massive Data Access Module
4. File Generation Module
Fault-Tolerance Mechanism
3.1 Operation Control Aspect

Operation Control Tool. Complete Operation Control Tools can improve software
usability, which include Operation Audit, Permission Management, and Maintenance
Management. Operation Audit audits software inputs and outputs and verifies that they
are legit. Permission Management authorizes different roles with corresponding per-
missions. Maintenance Management has two modules, including Flow Management
Module and Alert Notification Module. Flow Management Module manages sequence
of operations. Alert Notification Module notifies corresponding personnel with related
information, which can improve software availability.
3.2 Quality Control Aspect

Environmental Diagnostic Tool. Environmental Diagnostic Tool can help developers
focus on software functionality and business logic when defect occurs without being
interfered by systematic and environmental causes. Environmental Diagnostic Tool
includes database mapping check, configuration check, environmental IO check, etc.
Development Procedure & Checklist. Development Procedure can improve main-
tainability, readability, and reliability of software, and Checklist can help developers
make sure that they follow the Procedure. Development Procedure includes the naming
rule of functions and variables that helps improving maintainability, operation tips and
information that helps improving usability, and rules for configuration that helps
portability.
3.3 Software Functionality Aspect

Regression Test Tool. Regression Test Tool as in [9] can improve software accuracy
by minimize the effect to developed functions while developing new functions.
Components and Modules. With Components and Modules, we can aggregate sim-
ilar functions to improve development efficiency and shorten development time.
Data Group Control Module. Data Group Control Module can improve software
reliability by avoiding interference caused by concurrent executed operations of the
same group. Flexibility is achieved by adopting Aspect-Oriented Programming, AOP
as in [10, 11] concepts to add/remove functions flexibility.
Batch and Parallel Processing Module. Batch and Parallel Processing Module can
improve software flexibility by handling all issues related to batch and parallel pro-
cessing, such as parallelization, grouping mechanism, flow control, error handling, etc.
New functions can be developed by simply configuring settings or implementing
interfaces, and software performance can be tuned according to system resources
dynamically.
Massive Data Access Module. Massive Data Access Module can improve software
flexibility by de-coupling data access and business process. Therefore database or data
access tool can be upgraded or replaced without affecting software functionality.
File Generation Module. File Generation Module can improve software reusability by
de-coupling report format and data to achieve generating reports by setting data logic
and report format, which makes develop new reports/data-exchange files and modify
existing reports easier.
Fault-Tolerance Mechanism. Fault-Tolerance Mechanism can guarantee reliability
and accuracy of software by providing services stably and persistently. There are
hardware and software Fault-Tolerance Mechanism. Hardware Fault-Tolerance
Mechanism recovers software to latest backup when serious anomaly occurs. Soft-
ware Fault-Tolerance Mechanism includes solutions for pre-considered error-prone
situation and procedures for checking hardware/software errors.
4 Built-in Design Effectiveness
This study illustrates the improvement of various indicators by considering potential

failures and introducing the Built-in Design into software architecture in the early
stages of software development.
This study conducted experiments on two sets of software that fulfills the same
functional specifications on identical hardware equipment. One of the two sets of
software implements the Built-in Design functions this study designed in the early
stages of software development, and the other software does not. We inputted 10 sets of
test data which contains about 4.8 million test data in total to each software and check
the amount of incorrect results as shown in Fig. 2 as failure number. The failure
number of the software without Built-in Design is higher which averages 2700 and
varies dramatically among 10 sets of test data. On the contrary, the failure number of
the software with Built-in Design is 90% less which averages 200 and is stable among
10 sets of test data.
Fig. 2. Failures number of software with/without built-in design
We illustrate the Cost of Quality diagram of the two sets of software as shown in
Fig. 3 according to the method of measuring Cost of Quality, CoQ as in [12] and as in
(1), and a and b are coefficients of the project.
Cost of Quality ¼ a * Cost of Failure þ b * Cost of Prevention: ð1Þ
This study shows the results that under the same Cost of Failure, there is a sig-
nificant difference of Cost of Quality between software with Built-in Design and
software without it. The software with Built-in Design consumes lower unit Cost of
Prevention and results in about 64% reduction of the sum of the Cost of Quality
relatively.
Due to limited R&D resources, the Built-in Design functions based on FMEA was
designed in advance and implemented in three phases. At completion of each phase the
test data sets are inputted and output results are checked. The cost of checking failure
and removing defect are converted into Cost of Failure to obtain the Cost of Failure
curve as shown in Fig. 3.
Fig. 3. Cost of quality model
The Cost of Prevention curve as shown on the left chart in Fig. 3 shows that the
Cost of Prevention increases dramatically as software quality being improved. The
quality of software was improved by increasing the coverage of test and audit which is
achieved by increasing the number of test cases or the number of audit items. On the
contrary, the Cost of Prevention curve as shown on the right chart in Fig. 3 shows that
the Cost of Prevention increases much slower than the Cost of Prevention curve on the
left chart as software quality being improved. This is achieved by checking the
rationality and legitimacy of the input data and designing appropriate fault-tolerant
correction or error-handling mechanism accordingly and by incorporating functions of
partial logic segment testing when designing operational functions. Therefore the cost
of verifying multi-data input or complex logic conditions and the cost of auditing
execution results can be reduced.
This study found that through the Built-in Design of auditing and environmental
diagnostic tools, failures can be detected early and eliminated accordingly. Therefore the
operation can be completed within the specified time limit and failures detected in a later
phase and a series of following correction measures can be avoided. By developing a
development procedure and introducing automated continuous-integration mechanism
as in [13], defects of analysis, design, and development can be discovered before being
delivered for testing and deploying, which can reduce the failure rate by 25%.
5 Conclusion
By implementing the software with Built-in Design, we can apply quality assurance
measures as in [14] in the software development stage to improve development effi-
ciency and remove defects efficiently and stabilize software quality. Therefore we can
effectively reduce the cost of quality control measures as in [14] and achieves the goal
of producing high-quality software at lower cost that meets the concept of continuous-
integration.
This study analyzes Built-in Design required to software and explained it in detail.
Facing the diversified and endless software requirements, it is our goal to maintain the
quality and flexibility of software while satisfying user-requirements fast and
efficiently.
As Information Technology continues evolving, we will apply artificial intelligence
technology in the future to analyze and design the Built-in Design that can automati-
cally adapt the failure mode of environmental resources.
References
1. Markopoulos, P., Martens, J.-B., Malins, J., Coninx, K., Liapis, A.: Collaboration in creative
design: methods and tools, January 2016
2. McDermott, R.E., Mikulak, R.J., Beauregard, M.: The Basics of FMEA, 2nd edn.
Productivity Press, New York (2008)
3. Schultz, D.J.: A comparison of five approaches to software development, January 2000
4. Bourque, P., Fairley, R.E.D.: Guide to the Software Engineering Body of Knowledge. IEEE
Computer Society, New York (2014)
5. Cavano, J.P., McCall, J.A.: A framework for the measurement of software quality.
ACM SIGSOFT Softw. Eng. Notes 7, 133–139 (1978)
6. TM Forum, AI: The time is now, December 2017
7. Project Management Institute, A Guide to the Project Management Body of Knowledge:
PMBOK Guide, 5th edn., Project Management Inst. (2008)
8. Pentti, H., Atte, H.: Failure mode and effects analysis of software-based automation systems,
August 2002
9. Duggal, G., Suri, B.: Understanding regression testing techniques. In: National Conference
on Challenges and Opportunities. Mandi Gobindgarh, Punjab, India, 29 March 2008.
Accessed 3 Jan 2018
10. Kiczales, G., Lamping, J., Mendhekar, A., Maeda, C., Lopes, C., Loingtier, J.-M., Irwin, J.:
Aspect-oriented programming. In: Proceedings of the European Conference on Object-
Oriented Programming (ECOOP), June 1997
11. Laddad, R.: AspectJ in Action, 2nd edn. Manning, Greenwich (2009)
12. Campanella, J.: Principles of quality costs, 3rd edn. American Society for Quality Control,
Milwaukee (1990)
13. Humble, J., Farley, D.: Continuous Delivery: Reliable Software Releases Through Build,
Test, and Deployment Automation, 1st edn. Addison-Wesley, New York (2010)
14. Rose, K.H.: Project Quality Management: Why, What and How, p. 41. J. Ross Publishing,
Fort Lauderdale (2005). ISBN 1-932159-48-7
A Novel Sharing M-Coupons with Lightweight
Computations via Cloud Computing
Iuon-Chang Lin(&) and Hsiao-Chi Chiang
Department of Management Information Systems,

National Chung Hsing University, Taichung, Taiwan, Republic of China
iclin@dragon.nchu.edu.tw
Abstract. Coupons are useful and widely used in the market place to attract
new customers and keep loyal customers. As more and more wireless devices
are connected to the Internet, many have computed that mobile commerce will
blossom soon, just like e-commerce booming in recent years. In this paper, we
proposed a secure sharing m-coupon scheme that a user can use the mobile
device to share the sharable m-coupons to his/her friend through the cloud. In
the proposed scheme, a customer does not need to download the m-coupon from
cloud. If customer wants to redeem the m-coupon, he just needs to send a
redemption request message to the cloud, and the cloud will help the customer to
transfer the m-coupon to the retiler. So that the customers do not worry about
problem of storing and losing, it is convenient for customers. Our implemen-
tation incorporates only lightweight computations, message authentication code,
and simple hash chains for security.
1 Introduction
Nowadays, electronic commerce is very flourishing over the internet. Hence, many
traditional transactions have been proposed to perform electronically. As more and
more wireless devices are connected to the Internet, many have computed that mobile
commerce will blossom soon, just like e-commerce booming in recent years. In
addition, many marketing activities can be conducted on the mobile platform, such as
Short Message Service (SMS) advertising [1]. The mobile marketing is growing fast
because it is more cost effective, personalized, and real time delivery capabilities.
Therefore, advertisers develop digital coupons on mobile devices that make more
attract customers’ attention than traditional coupons. Juniper Research forecasted that
200 million mobile subscribers globally will use mobile coupons by 2013 [2].
A mobile coupon is defined as “a coupon sent to and stored on a mobile phone that
can be exchanged for a rebate, a monetary discount, etc. at a retailer or merchant when
the customer purchases a product [3].” A mobile user obtains an m-coupon in several
ways, such as downloading from a web site, an SMS message, a Multimedia Message
Service (MMS), or receiving it from a marketer or friend. Experts predicted a higher
usage of mobile coupons, which are stored in the memory of mobile device and
therefore difficult to misplace or forget, compared to their paper-based equivalents [4].
Therefore, the marketer can connect to their target customers more direct through the
mobile devices.

https://doi.org/10.1007/978-3-030-15035-8_43
460 I.-C. Lin and H.-C. Chiang
However, in the electronic environment, message can be easily intercepted and

tempered. Hence, how to protect the transaction message and build a secure channel for
m-coupons is a key issue. A pointed out in [5], every manufacturer coupon will be
digitally be the issuer. This is had to ensure that the three key field of a coupon: (A) the
product, (B) coupon value, and (C) the expiration date are not tampered with by either
the customer or the retailer. In addition, there are some security requirements for using
m-coupons [6, 7].
A. Authentication: The retailer should be able to ascertain whether the m-coupon has
been validly issued by an issuer. Thus, a retailer must have the ability to authen-
ticate the validity of all received m-coupons.
B. Integrity: The retailer should be able to detect whether the m-coupon data must
keep integrating. Thus, a retailer must have the ability to ensure that received m-
coupons have not been modified.
C. Preventing double spending: The number of m-coupons issued should be limited to
avoid unlimited use of m-coupons. Thus, the system must have the ability to
prevent an m-coupon be used twice.
In addition, in order to increase sales rate and attract new customers, manufacturers
may wish that a customer can share an unused m-coupon to another potential customer.
Otherwise, in order to keep loyal customers, manufacturers may issue special m-
coupons to loyal customers. These special m-coupons usually have more discounts.
Therefore, he may not wish that the special m-coupons be used by another customer.
There are many m-coupon schemes have been proposed [8, 9], motivated by sharing
secure m-coupons system [10], we attempt to issue two kinds of m-coupons and utilize
cloud to distribute m-coupons, one for royalty customer and the other for sharable
coupons. Hence, the customer can share the sharable m-coupons to his/her friends.
Therefore, we proposed a sharing m-coupon scheme, which may increase the sale rate
and benefit customers.
2 Related Work
2.1 Hash Chains

In our proposed scheme, we use the one- way and collision-resistance hash function
[11–14] to compute a hash chain. An e-coupon chooses a random number X_p and
computes the numbers X_(p − 1), X_(p − 2),…, X_0, where X_i = H(X_(i + 1)) for
i = p − 1, p − 2,…, 0. An e-coupon includes the hash values from X_0 to X_p, and
the X_0 is the root of the hash chain. When someone receives X_i, he can computes the
hash function i times with X_i to obtain X_0, where X_0 = H_i (X_i). The key feature
of hash chain is that if a value X_i is provided, it is not possible to find another X_j,
where j > i. Therefore, in order to verify X_i, the output of any hash value X_i can be
hashed i times and compared with the confirmation code X_0.
A Novel Sharing M-Coupons with Lightweight Computations via Cloud Computing 461
2.2 Message Authentication Code

Message authentication code (MAC) is a useful cryptographic technology that will be
helpful to our proposed scheme. The message authentication code is a special one-way
hash function [15–17] and provides such mechanism of integrity checks.
Typically, this algorithm accepts a secret key K and a message M as input, and it
then produces a message digest as output. It has the same features as the one-way hash
function, but it is expanded with a key to prevent someone from forging the hash value.
Only the user who owns the secret key can easily verify the hash function. If the
calculated MAC equals the received value, then the message is valid. This process
provides both an integrity check by ensuring that if message has been forged, and it is
an efficient cryptographic technology with lower operation cost.
3 Architecture of Proposed Scheme
Traditionally, one coupon is only sent to one user. To increase the redemption rate of
coupons, customer can share the coupons to his friends. Figure 1 shows the protocol of
this paper, which are described in the following. The proposed electronic coupon
scheme consists of six entities as follows:
Manufactures: The manufacturer who produces the goods, sells it to customers, and
authorizes the m-coupon provider to issue m-coupons.
Fig. 1. The architecture of electronic coupons.
Retailers: The retailer is the owner of the shops selling the promoted products. The
customers buy the promoted products at their online store. The retailer holds an
electronic point-of-sale device which is issued by the manufacture. A tamper-resistant
chip is embedded in to the device and contains the coupon provider’s secret key and the
verification function, such as a smart card or a SIM card. The chip in the device
protects the information stored in a secret area from copy or disclosing [18, 19].
Coupon provider: The coupon provider is the entity that issuing m-coupons to
potential customers through a wireless network. The coupon provider can accept a
registration request from a customer or a friend who has received a sharable m-coupon
from a customer.
Cloud: The cloud is the entity that distributing the m-coupons, which are issued by
the coupon provider, to potential customers or a friend who has received a sharable m-
coupon from a customer through a wireless network.
Customers: The customer is the client that buying the products at the retailer’s on-
line store and having registered in the coupon provider. A customer will receive both
m-coupons and sharable m-coupons from the coupon provider.
Friends: The friend will receive an authentication from the customer. He/She can use
the authentication to register with coupon provider to obtain a sharable m-coupon.
4 Proposed Scheme
4.1 Notations
Before proceeding to our scheme, we first introduce some notations that will be used
throughout the paper in describing the e-coupon system (Table 1).
Table 1. Notations for protocol description.

Notations Description
P Coupon provider
L Cloud
C Customer
F The friend of customer
R Retailer
M Manufacturer
IDC =IDF Unique identity of customer/friend such as mobile phone number
IDL =IDR Identity of cloud/retailer
MAC(K,M) The function for message authentication code. The inputs of this function are
a message M and a key K. The output of this function is a fixed-length
message digest
KP =KL =KC The coupon provider/cloud/customer’s secret key
PRODUCT The information on discounted products, such as product’s name and ID
number
SN Serial number for e-coupon
EXD Expiration of the e-coupon
PKR The public key of retailer
SKR The secret key of retailer
EKL =EKC =EKF Encrypted by the cloud/customer/friend’s secret key
(continued)
Table 1. (continued)
Notations Description
p The amount of e-coupons which customer wants to request
q The amount of sharable e-coupons which customer wants to request
Xi =Yj =Nl Position of the Chain X=Chain Y=Chain N
(i ¼ 0 p; j ¼ 0 q; l ¼ 0 q)
H() One-way hash function
∥ The concatenation of two strings
4.2 The Registration Phase

Before issuing the e-coupon, the cloud and customer have to register in the coupon
provider, and then the coupon provider will generate the secret keys for them,
respectively. This phase is performed only once.
Cloud L $ Coupon Provider P
Step 1: The cloud L sends a registration request and his IDL to the coupon provider P.
Step 2: After receiving the message from the cloud L, the coupon provider P uses
his secret key KP and the identity of cloud IDL to compute a secret key of cloud
KL ¼ MACðKP ; IDL Þ. He then sends KL to the cloud L through a secure channel.
Customer C $ Coupon Provider P

Step 1: The customer C sends a registration request and his IDC to the coupon
provider P.
Step 2: After receiving the message from the customer C, the coupon provider P
uses the cloud’s secret key KL and the identity of customer IDC to compute a secret
key of customer KC ¼ MACðKL ; IDC Þ. He then sends KC to the customer C
through a secure channel.
4.3 The Issuing Phase

In this phase, customer C asks a e-coupon request to the coupon provider P. Coupon
provider P generates e-coupons for customer, and he then transfers to the cloud L and
stores in there. In addition, coupon provider P transfers an authentication to the cus-
tomer C. If customer wants to share the e-coupons to his friends, he can give the
authentication to them, and they can use the authentication to withdraw the e-coupons.
Step 1: The C sends his identity IDC , the identity of cloud IDL , and the PRODUCT,
which he wants to buy with the discounted price, to the P.
Step 2: After receiving the message from C, P generates Chain X, ChainY , and
Chain N. This involves the following steps:
(1) Firstly, the P uses his secret key KP and IDL , which received from the C to
compute RL ¼ MACðMACðKP ; IDL Þ.
(2) The P uses the cloud’s secret key KL and IDC , which received from the C to
compute RC ¼ MACðMACðKL ; IDC Þ.
(3) The P chooses a random number Xp and computes the root value

X0 ¼ Hp Xp .
(4) The P chooses a random number Yq and computes the root value

Y0 ¼ Hq Yq .
(5) The P chooses a random number Nq and computes the root value

N0 ¼ Hq Nq .
(6) The P then use his secret key KP ; PRODUCT; SN; EXD; X0 RL ; RC to compute
SX ¼ MACðKP ; PRODUCT k SN k EXD k X0 k RL k RC Þ.
(7) The P then use his secret key KP ; PRODUCT; SN; EXD; Y0 RL ; N0 to compute
SY ¼ MACðKP ; PRODUCT k SN k EXD k Y0 k RL k N0 Þ.

Step 3: The P sends EKC Nq to C, which is encrypted by C′s secret key.

Step4: The P sends EKL IDC ; ðPRODUCT; SN; EXDÞ; Xp ; SX ; Yq ; SY ; N0 to
L, which is encrypted by L′s secret key.
Step 5: Upon receiving the message from P, the L decrypts the message by using his
secret key KL . And then he stores the value which in the message in his database.
4.4 The Redemption Phase for Customer

In this phase, the customer C wants to redeem his e-coupon, he will send a request
message to ask the cloud L, then the cloud L forwards the e-coupon to the retailer R,
who is the customer wants to purchase.
Step 1: The C using his secret key KC and identity IDC to computes
RC ¼ MACðKC ; IDC Þ. He then sends his identity IDC and EKC ððPRODUCT;
SN; EXDÞ; IDR ; RC Þ to the L, which is encrypted by C's secret key, where IDR is the
identity of retailer which C wants to purchase.
Step 2: After receiving the message from C, L acts the following steps.
(1) The L uses his secret key KL and the received identity of C to compute KC , and
then he decrypts the received message by KC .
(2) After that, the L uses his secret key KL and received IDC to computes RC ¼
MACðMACðKL ; IDC Þ; IDC Þ and checks whether it is equal to RC which he
received. If it is not, L stops the phase.
(3) The L computes the hash function p times with Xp to obtain X0 , where

X0 ¼ Hp Xp , and generate Xi .
(4) The L uses his secret key KL and identity IDL to compute RL ¼ MAC
ðKL ; IDL Þ.
(5) Finally, the L encrypts ððPRODUCT; SN; EXDÞ; Xi ; i; SX ; RC ; RL Þ using R′s
public key PKR to ensure that only the R can obtain the message, then he send it
to the R.
Step 3: After R receiving the redemption request message from L, he acts the
following steps.
(1) The R uses his secret key SKR to decrypt the message, and then checks whether
the expiration date of the e-coupon is valid.
(2) The R uses Xi and i, which are received from L, to computes the hash value
X0 ¼ Hi ðXi Þ.
(3) The R verifies whether the relation SX ¼ MACðKP ; PRODUCT k SN k EXD
k X0 k RL k RC Þ holds or not, where PRODUCT; SN; EXD; RC ; RL are
received from L, and X0 is computed by R in last step.
Step 4: If the above checks are correct, the R transfers PRODUCT; SN; EXD, Xi , and i
to M to see if it has been used before. Then the M computes the hash value X0 ¼
Hi ðXi Þ and finds the record of the e-coupon according to SN and checks whether the
X0 is correct and Xi has been used before. If the check indicates the m-coupon is valid,
the firm will send an acceptance message to the R and record the e-coupon as used.
Otherwise, if the check indicates the e-coupon is invalid, it means the m-coupon has
been used more than once. Then the M sends a rejection message to the R.
4.5 The Sharing Phase

In the issuing phase, the coupon provider issues m-coupons that include both the
Chain X and Chain Y. The former is used by the customer C, and the latter is used to
share the m-coupon with the friend F of customer. In this phase, we focus on sharing
m-coupons to the friend F of customer C.
Step 1: If F wants to request an e-coupon owning to C, the F has to send his identity
IDF to C.
Step 2: After receiving the message from F, the C checks whether F is his friend. If
so, C sends ðPRODUCT; SN; EXDÞ; Nl and l to F through a secure channel.
Step 3: Before using the e-coupon, the F has to register in the P. Hence, the F
forwards the received message from F with IDC to P.
Step 4: After receiving the message from F, P computes the hash function l times
with Nl to obtain N0 , where N0 ¼ Hl ðNl Þ. Then he uses L′s secret key KL , IDF and
N0 to compute the F′s secret key KF ¼ MACðKL ; IDF k N0 Þ, and he sends it to F.
When F wants to use the m-coupon, he just needs to send a redemption request
message to the cloud. The redemption phase is in the next phase.
4.6 The Redemption Phase for Sharable M-Coupons

In this phase, the customer F wants to redeem his e-coupon, he will send a request
message to ask the cloud L, then the cloud L forwards the e-coupon to the retailer R,
who is the customer wants to purchase.
Step 1: The F uses hid secret key KF and identity IDF to compute RF ¼ MACðKF ;
IDF Þ. He then sends ðPRODUCT; SN; EXDÞ; IDC ; IDF ; IDR ; Nl ; l, and RF to the L,
where IDC is the original owner of the e-coupon, IDR is the identity of R who is F
wants to purchase, and Nl and l is the authentication which C authorize to him in the
sharing phase.
Step 2: After receiving the message from F, L acts the following steps.
(1) The L computes the hash function l times with Nl to obtain N0 , where
N0 ¼ Hl ðNl Þ.
(2) After that, the L uses his secret key KL , received IDF , and N0 to computes
RF ¼ MACðMACðKL ; IDF k N0 Þ; IDF Þ and checks whether it is equal to RF
which he received. If it is not, L stops the phase.
(3) The L computes the hash function q times with Yq to obtain Y0 , where

Y0 ¼ Hq Yq , and generate Yj .
(4) The L uses his secret key KL and identity IDL to compute RL ¼ MACðKL ; IDL Þ.
(5) Finally, the L encrypts ððPRODUCT; SN; EXDÞ; Yj ; j; SY ; RL ; N0 Þ using R′s
public key PKR to ensure that only the R can obtain the message, then he send it
to the R.
Step 3: After R receiving the redemption request message from L, he acts the
following steps.
(1) The R uses his secret key SKR to decrypt the message, and then checks whether
the expiration date of the e-coupon is valid.
(2) The R uses Yj and j, which are received from L, to computes the hash value

Y0 ¼ Hj Yj .
(3) The R verifies whether the relation SY ¼ MACðKP ; PRODUCT k SN k EXD k
Y0 k RL k N0 Þ holds or not, where PRODUCT; SN; EXD; RC ; RL are received
from L, and Y0 is computed by R in last step.
Step 4: If the above checks are correct, the R transfers PRODUCT; SN; EXD, Yj , and j
to M to see if it has been used before. Then the M computes the hash value Y0 ¼

Hj Yj and finds the record of the e-coupon according to SN and checks whether the
Y0 is correct and Yj has been used before. If the check indicates the e-coupon is valid,
the firm will send an acceptance message to the R and record the e-coupon as used.
Otherwise, if the check indicates the m-coupon is invalid, it means the e-coupon has
been used more than once. Then the M sends a rejection message to the R.
5 Security Analysis
5.1 Authentication
Our proposed scheme achieves this requirement by allowing a retailer to make sure the
validity of e-coupons. In the e-coupon issuing phase of our scheme, a valid e-coupon
only generates by the genuine coupon provider with the secret key. Any other user in
this system cannot generate a valid m-coupon without the knowledge of the secret key
of the genuine coupon provider. In addition, if a customer attempt to forge an
m-coupon on behalf a coupon provider is impossible, because the forger difficult to get
valid SX or SY .
In the issuing phase, the coupon provider signs PRODUCT, SN, EXD, X0 (or Y0 ),
RL , and RC (or N0 ) by using the secret key of coupon provider to generate SX or SY ,
which is used for verification during redemption. If a malicious user attempts to redeem
a forged m-coupon, the verification will fail. Because the forged m-coupon would have
been generating by using illegal versions Xp and Yq . In the redemption phase, when the
retailer received illegal Xi the retailer will compute the one-way hash function i times
and compare the output with the root value X0 in the SX . Therefore, a retailer can detect
the validity of m-coupon by SX (or SY Þ in this system. In addition, in the redemption
phase for sharable m-coupons, if an attacker attempts to intercept the message to obtain
the Nl and l, he cannot succeed. Due to the Nl and l just can be used for only once and
the cloud will check it with the identity of friend, so the entity of friend does not worry
about they may be intercepted or copy by attackers.
Furthermore, a retailer also can verify whether the m-coupon received from an
assigned customer. In the redemption phase, the RC (or RF ) is generated from customer
by using his/her secret key KC (or KF ) and identity IDC (or IDF ), and transfer it to the
cloud. When cloud received the RC (or RF ), he can verify it by using his secret key KL
and IDC . We can assume the secret key KC (or KF ) and KL are issued by a trust party.
Since the secret key KC ¼ MACðKL ; IDC Þ (or KF ¼ MACðKL ; IDF ) is generated with
the secret key of KL , so that the corresponding cloud can verify the validity of identity
of customer. In addition, when the retailer received the redemption message from the
cloud, he also can verify the identity of cloud and customer through SX (or SY Þ with the
KP , which is stored in the point-of-sale device. Therefore, our proposed scheme can
detect any forged or false m-coupon successfully.
5.2 Integrity
In the electronic environment, message can be modified easily especially in the
transaction processes. Therefore, when a retailer receives an m-coupon and its asso-
ciated S_X (or S_Y), we must to ensure the received m-coupon has not been modified.
We assume that a malicious customer or cloud attempts to modify PRODUCT to
change the product information or modify EXD to extend the expiration date. The
modification will not pass the verification of S_X (or S_Y) in the redemption phase.
Therefore, our proposed scheme also can verify the integrity of received m-coupons.
5.3 Single Use

The number of m-coupons issued should be limited to avoid unlimited use of m-
coupons. Thus, the single use requirement must be satisfied. In our scheme, a retailer
can make sure that a received m-coupon is used only once by checking the manu-
facturer’s database. When the manufacturer received the message from retailer, he will
find the record of the m-coupon according to SN and checks whether the m-coupon has
been used before. Hence, if a customer uses the m-coupon more than once, the man-
ufacturer can easily detect the problem by using its database system. Therefore, our
proposed scheme also can prevent the double spending.
5.4 Computational Cost

In our proposed scheme, MAC and one-way hash function are the main operation for
generating and verifying m-coupons. So the number of MAC operations and one-way
hash function can be used to decide the computation cost. According to the ISO
standard of a MAC algorithm [16], the MAC algorithm encrypts the hash value, which
is computed using the MD5 hash function, by using DES encryption function.
Therefore, MAC’s computation cost is one hash function and one DES encryption
function computation. Table 2 shows the number of public key signature, symmetric
key encryption, and one way hash function that can be performed on a typical work-
station [20, 21]. According to the results, we find that the one-way hash function and
the symmetric key encryption are much cheaper than the public key signature.
Table 2. Comparisons of the computational speed.

Operation Number per second
One-way hash function (MD5/SHA) 20,000
Symmetric key encryption (DES) 2,000
Public key signature (1024 bits RSA) 2
Table 3 lists all of the operations in each phase except for the registration phase.
We can see that the proposed schemes indeed require very low computational cost.
Furthermore, the computational capability of the customer’s mobile phone is very
small. Therefore, our proposed scheme can reduce the load of customer’s mobile
phone, and it is practical in the mobile commerce.
Table 3. Analysis of the computational cost for the entities in the proposed scheme.
Phase Entity Operation Times
Registration Customer MAC 0
One-way hash 0
Encryption/Decryption 0
Issuing Customer MAC 0
One-way hash 0
Decryption 1
Redemption Customer MAC 1
One-way hash 0
Encryption 1
Sharing Customer MAC 0
One-way hash 0
Friend MAC 0
One-way hash 0
Redemption for Friend MAC 1
sharable m-coupons One-way hash 0
6 Conclusions
In this paper, we have presented a secure sharing m-coupon scheme that a user can use
the mobile device to share the sharable m-coupons to his/her friend through the cloud.
Our proposed scheme combines the popularity of cloud to improve coupon redemption
rate, it may benefit both customer and markets. In addition to achieving the three
security requirements- (1) authentication, (2) integrity, and (3) single use, the proposed
scheme also provides (1) low computational cost, and (2) low communication cost. It
can not only be used on the Internet, but also suitable for being used on wireless
networks. Therefore, the proposed scheme is very efficient and practical.
References
1. Scharl, A., Dickinger, A., Murphy, J.: Diffusion and success factors of mobile marketing.
Electron. Commer. Res. Appl. 4(2), 159–173 (2005)
2. Neil, D.S.: Coupons: Identifying New Opportunities Beyond Early Trials. Jupiter Research
Report (2008)
3. Juniper Research: Mobile Coupons (2008). http://www.juniperresearch.com
4. Raskino, M.: Mobile coupons will reach right into your pocket. GartnerGroup Research Note
(2001). http://www.gartner.com/
5. Kumar, M., Rangachari, A., Jhingran, A., Mohan, R.: Sales promotions on the internet. In:
Proceedings of the 3rd USENIX Workshop on Electronic Commerce, pp. 167–176 (1998)
6. Blundo, C., Cimato, S., De Bonis, A.: A lightweight protocol for the generation and
distribution of secure e-coupons. In: Proceedings of the 11th International Conference on
World Wide Web, Honolulu, Hawaii, USA, pp. 542–552 (2002)
7. Blundo, C., Cimato, S., De Bonis, A.: Secure E-coupons. E-Commer. Res. 5(1), 117–139
(2005)
8. Chang, C.C., Wu, C.C., Lin, I.C.: A secure E-coupon system for mobile users. Int.
J. Comput. Sci. Netw. Secur. 6(1), 273–279 (2006)
9. Wray, J., Plante, D.: Mobile advertising engine for centralized mobile coupon delivery. Int.
J. Manag. Mark. Res. 4(1), 75–85 (2011)
10. Hsueh, S.C., Chen, J.M.: Sharing secure m-coupons for peer-generated targeting via eWOM
communications. Electron. Commer. Res. Appl. 9(4), 283–293 (2010)
11. Ricest, R.L., Shamir, A.: PayWord and MicroMint: two simple micropayment schemes. In:
Proceedings of 1996 International Workshop on Security Protocols, pp. 69–87 (1996)
12. Damgard, I.B.: A design principle for hash functions. In: Advances in Cryptology-CRYPTO
1989. LNCS, vol. 435, pp. 416–427 (1990)
13. Merkle, R.C.: One way hash function and DES. In: Advances in Cryptology-CRYPTO
1989. LNCS, vol. 435, pp. 428–446 (1989)
14. Merkle, R.C.: A fast software one-way hash function. J. Cryptol. 3(1), 43–58 (1990)
15. Davies, D.W.: A message authentication algorithm suitable for a main-frame computer. In:
Advance in Cryptology-CRYPTO 1984. LNCS, vol. 196, pp. 393–400 (1984)
16. ISO/IEC 9797: Data cryptographic techniques-data integrity mechanism using a crypto-
graphic check function employing a block cipher algorithm. Internal Organization for
Standardization
17. Menezes, A., Van Oorschot, P., Vanstone, S.: Handbook of Applied Cryptography. CRC
Press Inc., Boca Raton (1996)
18. Hendry, M.: Smart Card Security and Applications. Artech House, Norwood (1997)
19. Jones, H.W.E., Watson, A.C., O’Neill, T.J.: Vehicle security using smartcards. Secur. J.
13(2), 79–87 (1998)
20. Schneier, B.: Applied Cryptography, 2nd edn. Wiley, Hoboken (1996)
21. Baldonado, M., Chang, C.-C.K., Gravano, L., Paepcke, A.: The stanford digital library
metadata architecture. Int. J. Digit. Libr. 1, 108–121 (1997)
A Smart Roadside Parking System Using
Bluetooth Low Energy Beacons
Hui-Tzu Chen, Pei-Yu Lin, and Chi-Yi Lin(&)
Tamkang University, Taipei, Taiwan, R.O.C.

chiyilin@mail.tku.edu.tw
Abstract. The management of roadside parking has been a big issue for
metropolitan areas. Many city governments still rely on manual issuing of
parking tickets or use parking meters to collect the parking fee. Either way, it is
not economical in terms of the expenses on human labor and the efficiency of
parking space management. To solve the problem, some researchers have pro-
posed roadside parking systems based on wireless sensor networks. However,
the existing approaches focus only on detecting the occupancy of parking
spaces. This is not enough for a comprehensive roadside parking system because
the identification of vehicles using specific roadside parking spaces is yet
unknown. In this work-in-progress research, we propose a roadside parking
system based on the Bluetooth Low Energy (BLE) beacon technology to
achieve the two goals simultaneously. Specifically, vehicles are attached with a
BLE beacon device which broadcasts the identification of the registered vehicle
to be scanned by the roadside units. By monitoring the estimated distances
between the beacon device and a number of roadside units, our backend system
is expected to determine the occupancy of the roadside parking spaces along
with the identification of vehicles using the parking spaces.
Keywords: BLE beacon RSSI Roadside parking
1 Introduction
With the recent advancement of information and communications technologies, many

city governments around the globe have commenced to deploy ICT infrastructure for
the purpose of building smart cities. Smarter public safety, energy, buildings, trans-
portation, and water management are the key application scenarios in the context of
smart cities. Let’s take the smart street lights of the Smart Taipei project [1] as an
example. The street lights in the demonstration field are equipped with environmental
sensors and cameras to calculate the traffic flow and crowds, and to detect the occu-
pancy of roadside parking spaces. With 5G/NB-IoT interfaces, these smart street lights
are able to serve as a common carrier for a variety of environmental sensor data, so as
to promote innovative and smart applications. In this research, we focus on the topic of
roadside parking, which is one of the key issues in the field of smart transportation
management.
The importance of making roadside parking management smarter is of no doubt.
Traditional roadside parking management requires manpower to go around and inspect

https://doi.org/10.1007/978-3-030-15035-8_44
472 H.-T. Chen et al.
the parking spaces regularly, and then issue parking tickets accordingly. Obviously,
this traditional approach may consume a large amount of human labor and time. To
reduce the manpower needs, many city governments use parking meters to collect
parking fees. Although this approach automates the process of fee collection, it only
solves part of the problem in roadside parking management – the management
authority is still unable to know the real-time status (occupied or not) of each parking
space. For a driver looking for an unoccupied parking space, the driver needs drive
along the street to identify an available one. This is inefficient in terms of time spent
looking for a parking space, and introduces unnecessary carbon emissions.
With the latest ICT technologies, many researchers proposed novel approaches to
build a smart roadside parking management system. For instance, in 2013 Zhang et al.
[2] proposed a street parking system based on wireless sensor networks. They deployed
magnetic sensor nodes on every parking space and designed a vehicle detection
algorithm to detect a parking car. In 2014 Barone et al. [3] proposed the intelligent
parking assistant (IPA) architecture where RFID readers and inductive magnetic loops
are deployed on every parking space to detect whether the space is available or
occupied. They even allowed drivers to reserve a parking space at a specific destination
area before the driver departs for the destination. However, we found that with these
approaches, the management system can merely monitor the status of the roadside
parking spaces. More specifically, the management system has no idea of the identi-
fication of the vehicles using the parking spaces, which makes it difficult to bill the
vehicle owners automatically. To solve this problem, some vendors have offered their
solutions such as the smart parking meter developed by AcerITS [4], a Taiwan-based
company. The smart parking meter has a built-in camera which can recognize the
license plate and then uses it for billing. However, license plate recognition cannot
achieve a perfect accuracy of vehicle identification under poor illumination conditions
or obstruction of license plates.
Based on the above observations, in this research we would like to take advantage
of a popular and low-cost technology – Bluetooth Low Energy (BLE) beacons, to build
a smart roadside parking management system. As long as each vehicle is attached with
a registered beacon device, we expect that our system can detect the occupancy of
every roadside parking space and accurately acquire the identification of the vehicle
occupying a specific parking space. Note that this is a work-in-progress research, in
which we are still working on the problem of distance estimation based on radio
signals.
The rest of this paper is organized as follows. In Sect. 2 we introduce the back-
ground knowledge for the Bluetooth beacon technology. In Sect. 3 the system archi-
tecture along with the principle of operations are described. Section 4 shows the
prototype implementation, which includes how we acquire the identification of vehicles
and how to detect the occupancy of a roadside parking space, as well as how we deploy
the Bluetooth network to support the application-layer protocol – MQTT. Finally, the
conclusions and future work are given in Sect. 5.
A Smart Roadside Parking System Using BLE Beacons 473
2 Background Knowledge
In this section we will introduce the Bluetooth beacon technology. According to the
Bluetooth Core Specification version 4.1, the maximum size of a BLE advertising
packet data unit (PDU) is 39 bytes, in which the advertising data is up to 31 bytes. The
advertising data field is exactly where various beacon protocols used to carry the
beacon data. Although the latest Bluetooth specification (version 5.0) has extended the
maximum advertising channel PDU size to 257 bytes, in Bluetooth 5.0 the so called
“legacy advertising PDUs” are still up to 39-byte long in order to make them backward
compatible.
There are a number of Bluetooth beacon pseudo-standards in the market, among
which the most popular ones are Apple’s iBeacon [5], Google’s Eddystone [6], and
Radius Networks’ AltBeacon [7]. As shown in Fig. 1(a), the iBeacon packet format
consists of four fields: 16-byte UUID, 2-byte Major, 2-byte Minor, and 1-byte Tx
Power, where UUID stands for Universally Unique Identifier. According to [5], a
UUID should be specific to an application/deployment use case, and Major/Minor
values are used to hierarchically specify sub-regions within a larger region or use case.
As for the meaning of Tx Power, it is defined as the wireless signal strength at a
distance of 1 m away from the beacon transmitter. By comparing the received signal
strength indicator (RSSI) and the value of Tx power embedded in the beacon packet, a
beacon receiver can estimate its proximity to the beacon transmitter.
Fig. 1 Packet formats of three popular Bluetooth beacons: (a) iBeacon, (b) Eddystone-UID, and
(c) AltBeacon.
The packet formats of Eddystone-UID and AltBeacon are shown in Fig. 1(b) and (c),
respectively. Note that Eddystone actually defines four packet types: Eddystone-UID,
Eddystone-EID, Eddystone-URL, and Eddystone-TLM. Since we focus on identifying
vehicles by UUIDs, we only introduce Eddystone-UID here. In Eddystone-UID we see
the 10-byte Namespace and the 6-byte Instance fields, which can be combined to form a
16-byte universally unique beacon ID. Google recommended two methods to generate a
unique 10-byte namespace: a truncated hash of owner’s FQDN, or an elided UUID [8].
The field of Ranging Data is designed to carry the calibrated Tx power at 0 m, which can
be obtained by measuring the actual signal strength from 1 m away and then add
41 dBm to the value. As for AltBeacon, we can see the 20-byte Beacon ID and the 1-
byte Ref RSSI fields. For interoperability purposes, the AltBeacon specification indicates
that the first 16 bytes of the Beacon ID should be unique to the advertiser’s organiza-
tional unit. The Ref RSSI field is used to represent the average received signal strength at
1 m from the advertiser, which is of the same definition as the Tx power field in iBeacon.
To sum up, the above three types of beacon packets serve a common purpose: carrying a
beacon ID as an identification and a transmit power value for proximity estimation.
3 System Design
3.1 System Architecture
The proposed system architecture is depicted in Fig. 2. There are mainly four roles in
our system: beacon transmitter, beacon reader, gateway, and backend server. We
assume that each vehicle has a registered beacon transmitter attached to the right side
mirror (for right-hand traffic countries). The beacon transmitter emits beacon packets
regularly, which then can be detected by the beacon readers deployed at every corner of
the parking spaces along the curb. A gateway device playing the Bluetooth central role
maintains a connection to each beacon reader playing the Bluetooth peripheral role. In
order to run application-layer protocols in the Bluetooth network, we need the gateway
device and the beacon readers capable of running the IP protocol. As a result, we
deployed the 6LoWPAN technology in the Bluetooth network, where the gateway
serves as the 6LoWPAN Border Router (called 6LBR), and all the readers serve as the
6LoWPAN Nodes (called 6LN). Each reader uses the received beacon packets to
identify the vehicles and to estimate the distances between the reader and the nearby
vehicles. Finally, the readers send the vehicle IDs and the estimated distances to the
backend server for parking status analysis.
3.2 Principle of Operations

As we mentioned in Sect. 2, all iBeacon, Eddystone-UID, and AltBeacon packets
contain a UUID field. We assume that each vehicle is allocated with a UUID (named
vehicle ID hereafter), which then can be configured in the beacon transmitter of the
vehicle. Moreover, beacon packets also carry their transmit power values in the Tx
power field. When a beacon packet is received at the beacon reader, the reader can use
the RSSI of the beacon packet along with the Tx power value to estimate the rough
distance between a specific beacon transmitter and the reader. After that, every reader
can send its own list of three-tuples (vehicle ID, estimated distance, reader ID) to the
backend server. Since the same vehicle ID can be detected by a number of beacon
readers, the backend server may receive several tuples of the same vehicle ID
Fig. 2 System architecture for roadside parking management using BLE beacons.
associated with different reader IDs. To determine the parking space occupied by a
specific vehicle, the server can simply sort the tuples of the same vehicle ID by the
estimated distances. We may use the scenario in Fig. 2 to give an example, where the
vehicle is on the parking space with ID 02. Let’s assume the vehicle ID of this vehicle
is UUID1. All the four readers detected UUID1 and sent their estimated distances to the
backend server. After the server sorted all the tuples by the estimated distances (the
example shown in Table 1), the server found that the vehicle is closest to Reader2,
followed by Reader3. Meanwhile, the server further found that the top-2 distances
(namely, d1 and d2) are below some predetermined thresholds (e.g., d1 < 4 and d2 < 6).
Finally, the server can conclude that the vehicle is on the parking space between
Reader2 and Reader3 (i.e., the parking space 02).
Table 1 An example of sorted tuples for the vehicle with UUID1.

Vehicle ID Estimated distance (m) Reader ID
UUID1 2.15 Reader2
UUID1 4.33 Reader3
UUID1 9.09 Reader1
UUID1 11.42 Reader4
4 Prototype Implementation
In our prototype system, the gateway and the readers are implemented by using
Raspberry Pi 3 development boards which have built-in Bluetooth v4.1 interfaces, and
the beacon transmitter is implemented by using an Estimote Proximity Beacon device
which supports both iBeacon and Eddystone BLE beacon packets. Currently, we
configure the beacon device to emit iBeacon packets in order to test the functionality.
In the following we will describe the process of distance estimation carried out by
beacon readers, and the way we deploy the 6LoWPAN-based Bluetooth network.
4.1 Vehicle Identification and Distance Estimation

The primary goals of the beacon readers are to discover the identification of nearby
vehicles and to estimate the distances between the readers and the vehicles, both by
analyzing the received iBeacon packets. Therefore, the first thing we need to do is to
filter iBeacon packets out of all BLE air packets. To determine whether a BLE packet is
an iBeacon packet, we can look into the AD structures which constitutes the Bluetooth
advertising data. As shown in Fig. 3, an iBeacon packet is composed of two AD
structures, where the content of the 9-byte iBeacon Prefix field is fixed. By installing
the bluepy package [9] on the Raspbian OS, we can do blescan to scan for BLE
packets. The scanning result is shown in Fig. 4, where the highlighted part is exactly
the iBeacon packet advertised by our beacon transmitter with the MAC address
D3:F0:4E:47:B0:6A. Note the content of the Manufacturer field starts with
0x4c000215, which matches the values of the 4 bytes Company ID, SubType, and
SubType Length fields in AD2. For simplicity, we modified the source code of
blescan by adding a filter to match the string 0x4c000215. Therefore, we can
successfully capture the vehicle UUIDs by parsing the received iBeacon packets.
Fig. 3 Decomposition of iBeacon Prefix field.
To estimate the distance between a beacon reader and a beacon transmitter, we need
the following two pieces of information: the RSSI of the received beacon packet and the
Tx power designated in the beacon packet. In fact, both pieces of information are
available from the previous execution output. From Fig. 4 we can see the value
Fig. 4 Scanning results of BLE packets.
of −67 dBm, which is the RSSI of the detected iBeacon packet. The value of Tx power
in 2’s complement is 0xC5 (i.e., −59 dBm in decimal), which is located at the last byte
of the Manufacturer field. Then we implemented two different formulas to calculate the
estimated distance d. The first formula is from [10] as shown in (1):
ARSSI
d ¼ 10 10 n ð1Þ
Here n is the signal transmission constant which can be found by empirical

experiments (normally between 2 and 2.5), and A is the RSSI at 1 m. The second
formula is from the Android beacon library [11] as given in (2):
7:7095
RSSI
d ¼ 0:89976 þ 0:111 ð2Þ
A
Then, we modified the source code of blescan again by adding the two formulas
to do distance estimation. Figure 5 shows one of the experimental results, where
Distance_M1 and Distance_M2 refer to the distance estimation by using (1) and (2),
respectively. In fact, we have evaluated the accuracy of the distance estimation at 8
different distances ranging from 1 m to 8 m (incremented by 1 m), and at each distance
we collected the distance estimation for 20 times. However, we found that the errors of
both formulas are quite large due to the significant fluctuation of radio signals. We are
still working on this issue and would design a numerical algorithm to smooth the
collected data. As it has been pointed out in the official blog of the Bluetooth SIG [12],
the author suggested that one should avoid using the absolute value of the RSSI – use
the trend instead. Therefore, we will try to take advantage of the trend of the RSSI to
validate the applicability of our proposed approach.
Fig. 5 Output of distance estimation between the beacon transmitter and the reader.
4.2 Deployment of 6LoWPAN-Based BLE Network

For parking status analysis, the backend server must collect the distance estimation data
from the beacon readers. In our prototype system, a gateway device sits between the
backend server and the beacon readers, which is responsible for maintaining BLE
wireless links to each reader and then connects to the backend server. As we mentioned
in Sect. 3.1, we deployed the technology of 6LoWPAN over BLE in the Bluetooth
network consisting of the gateway and the readers. This is done by first establishing a
link-layer connection between a gateway and a reader using BlueZ [13], and then we
load the bluetooth_6lowpan module in Raspbian and enable the 6LoWPAN
functionality. After that, at the gateway (i.e., 6LBR) we use Linux IPv6 router
advertisement daemon (called radvd) to send the IPv6 Router Advertisement
(RA) messages for the readers (i.e., 6LNs) to generate their own global IPv6 addresses.
By now, all the devices in the 6LoWPAN network are addressable through a public
IPv6 address. Figure 6 shows the IPv6 configuration of the Bluetooth interface at one
of the 6LNs, as well as the result of a ping test from the 6LN to the Internet.
Fig. 6 IPv6 configuration of the Bluetooth interface and the ping test at a 6LN.
Next, we need to design the application architecture that can send the distance
estimation data back to the backend server in a timely manner. We choose the popular
application-layer protocol suitable for IoT environments – MQTT [14], to implement
the Publish/Subscribe messaging model. As shown in Fig. 7, beacon readers play the
role of publishers which publish a specific topic containing the distance estimation data
to the gateway acting as a broker. The backend server is configured to subscribe to this
specific topic. Whenever a new piece of distance estimation data is published by any
reader, it will be pushed to the backend server automatically. Consequently, the
backend server can successfully collect the needed data and then perform parking status
analysis. The whole MQTT-based messaging system is implemented using the open-
source Really Small Message Broker (RSMB) [15].
Fig. 7 Publish/Subscribe messaging model for our application architecture.
In this research, we proposed a smart roadside parking management system based on

the BLE beacon technology. Compared with the existing management systems, our
system can determine the occupancy of a roadside parking space and the identification
of the vehicle using the parking space simultaneously. Moreover, unlike the approaches
based on license plate recognition that suffer from either obstruction or illumination
issues, our system can detect the vehicle identification easily in all conditions, on a
minor cost of deploying a beacon device on each vehicle.
Currently our work-in-progress prototype system is merely a proof of concept.
Considering the problem of fluctuation of radio signals, in the near future we will try to
design a more sophisticated data processing algorithm in order to improve the stability
of parking status analysis. The scalability of the prototype system is limited by the star
topology of the 6LoWPAN-based BLE network. In the future, this problem can be
overcome by transitioning the network topology into a BLE mesh network. Moreover,
in terms of better user experience, a smartphone app would be necessary for vehicle
drivers to confirm their parking statuses. Last but not least, to prevent the problem of
counterfeit vehicle IDs, the beacon ID should be protected by some security
mechanisms.
References
1. Smart Taipei - Smart street light empirical field. https://smartcity.taipei/project/120
2. Zhang, Z., Li, X., Yuan, H., Yu, F.: A street parking system using wireless sensor networks.
Int. J. Distrib. Sens. Netw. 9, 1–10 (2013)
3. Barone, R.E., Giuffre, T., Siniscalchi, S.M., Morgano, M.A., Tesoriere, G.: Architecture for
parking management in smart cities. IET Intell. Transport Syst. 8, 445–452 (2014)
4. Acer ITS - On Street Parking - Smart Parking Meter. http://www.acerits.com/index-en.
html#works
5. Apple iBeacon. https://developer.apple.com/ibeacon/
6. Eddystone. https://developers.google.com/beacons/eddystone
7. AltBeacon. https://altbeacon.org/
8. Eddystone-UID. https://github.com/google/eddystone/tree/master/eddystone-uid
9. bluepy - Python interface to Bluetooth LE on Linux. https://github.com/IanHarvey/bluepy
10. Wang, Y., Xu, Y., Yutian, Z., Yue, L., Cuthbert, L.: Bluetooth positioning using RSSI and
triangulation methods. In: 2013 IEEE 10th Consumer Communications and Networking
Conference (CCNC), pp. 837–842 (2013)
11. AltBeacon Android Beacon Library. https://github.com/AltBeacon/android-beacon-library
12. Proximity and RSSI. https://blog.bluetooth.com/proximity-and-rssi
13. BlueZ - Official Linux Bluetooth protocol stack. http://www.bluez.org/
14. MQTT Version 3.1.1. http://docs.oasis-open.org/mqtt/mqtt/v3.1.1/os/mqtt-v3.1.1-os.html
15. RSMB: Really Small Message Broker. https://github.com/eclipse/mosquitto.rsmb
A Light Weight Stream Cypher Mechanism
for Visible Light Communication
Shih-Hao Chang1, Ted Huang2(&), and Mei-Lan Chen3

1
Department of Computer Science and Information Engineering,
Tamkang University, New Taipei City 25137, Taiwan
shhchang@mail.tku.edu.tw
2
Chunghwa Telecom Laboratories, Taipei City 10682, Taiwan
tshuang@cht.com.tw
3
Graduate School of Cultural Creative Design,
Chungyu University of Film and Arts, Keelung City 20103, Taiwan
katrinachen@cufa.edu.tw
Abstract. LED light has many advantages such as power efficient, better
quality, long lifespan, and focus emission. LED lighting has been employed in
our daily life in various ways, such as car lamps and lights in vehicles, roadside
signal lights, indoor lighting systems, etc. The LED light bulbs not only can
provide bright but also can be employed for data communication and object
positioning, namely visible light communication (VLC). VLC is fast-growing
technologies that enable new visible light application such as fast data trans-
mission and indoor positioning. Due to its attractive functionalities, they have
draw lots of attention to the wireless communication and indoor navigation
systems. Because of its physical characteristics, “what you see is what you
send”, VLC has been always considered as a secured communication method. In
this work, we analyze the risks of visible light communication signal such as
signal jamming, data snooping and modification. In order to effectively protect
VLC data communication problems, we design a lightweight stream cypher data
communication through the VLC signal to achieve the secure of the data
communication. Experimental results of the proposed mechanism show that the
WG-8 encryption algorithm can be executed on Raspberry Pi3 Model B, which
has limited computing power. The evaluation result shows WG-8 support
embedded system and CPU utilization reach up to 49%.
Keywords: Visible light communication Error propagation

Cypher security Stream cypher WG-8
1 Introduction
Using LED lighting has many advantages such as power efficient, better quality, long
lifespan, and focus emission. LED lighting has been employed in our daily life in
various ways, (car lamps and lights in vehicles, indoor lighting systems, etc.). Visible
Light Communication (VLC) is a fast-growing technology to provide data communi-
cation using low-cost and energy efficient light-emitting diode (LED) and photodiodes.
Due to its attractive functionalities, it has draws lots of attention to the wireless

https://doi.org/10.1007/978-3-030-15035-8_45
482 S.-H. Chang et al.
communication and navigation systems that also reduces cost after system employed.
This white visible light source can be modulated at high-speed so there is the potential
to transmit data as well as provide illumination, e.g., the visible light communications
consortium [1], the IEEE task group, 802.15.7 [2], standardizing VLC for personal area
network etc. Since history likes to repeat itself a common mistake in the development
of novel communications techniques was to neglect or downplay the security issues:
such was the case with the internet protocol suite, fiber-optics networks and etc.
One of the features in which VLC techniques are considered superior to traditional
wireless communication is security. High directivity and low impermeability of visible
light signal are considered to provide a secure way to transmit data between two peers,
making the data difficult to intercept from outside. The common slogan summarizing
VLC security features is: “What you see is what you send” – WYSIWYS [3]. In this
paper we address the classic security triad of: confidentiality, authenticity and integrity
in VLC. Although block cryptography is more widely used than streaming cryptog-
raphy, however, most of the visible light communication required in terms of time
urgent. Stream encryption is more instantaneous than block encryption, and consumes
less power. Because the data encryption by the block cypher will cause more resources
and increasing complexity. It is not suitable for the block cypher system in visible light
communication, and the stream encryption method for bit-based encryption is much
suitable for use in visible light communication.
To address this problem, we proposed a reliable VLC based data communication
mechanisms. First, link quality-monitoring agents, which will able to identify the
faulty, link which enabling receiver node to notify its transmitter node and resume
connectivity after a link breakage. Second, a reliable visible light transmission system
has been proposed, includes three major modules including environment monitoring,
intelligent control and high speed transmission. The rest of the chapter is organized as
follows. In Sect. 2, background and related works are briefly described. In Sect. 3, we
present our Link Guard Scheme (LGS) and its detail mechanisms. Finally, our con-
clusions and future work are explored in Sect. 5.
2 Background and Related Works
Visible Light Communication (VLC) is a fast-growing technology to provide data

communication using low-cost and energy efficient light-emitting diode (LED) and
photodiodes. For reasons of simplicity and cost most white light sources use Gallium
Nitride (GaN) emitters. The blue emission excites a yellow phosphor layer deposited
on top of the LED ‘chip’, and white light is created by the addition of the blue and
yellow radiation. This white visible light source can be modulated at high-speed so
there is the potential to transmit data as well as provide illumination which has drawn
interest from both research and industrial communities, e.g., the Visible Light Com-
munications Consortium [1], the IEEE task group, 802.15.7 [2], standardizing VLC for
personal area network etc.
In the VLC system, LOS (Line of Sight) communication between two VLC
transceivers constitutes the majority of the light communication applications. With this
leading edge technology, data including video and audio, internet traffic, etc., can be
A Light Weight Stream Cypher Mechanism for Visible Light Communication 483
transmitted at high speeds using LED light. However, since the link can cause link
broke due to movement such as object slight movement or rotation of the receiver unit.
The poor pointing of the VLC devices may cause the decrease of signal quality or even
link disconnection. Moreover, VLC can be highly directional, it is difficult to re-
establish link that has been lost due to movement or rotation of one of the devices in the
link [4–6]. As such, it is necessary and important to have link recovery as well as
handover mechanisms for VLC systems, yet the fast link recovery is very challenging.
Encryption [7] is a mathematical algorithm used in cyber security for data pro-
tection. Usually there are keys and texts and cipher texts. The interception is just a
cipher text that cannot be understood. When the receiver receives the cipher text, it is
transferred back to the text by the decryption algorithm. If the encryption and
decryption keys are the same, then called symmetric encryption algorithm [8, 9]. If
keys are different, is called an asymmetric encryption algorithm [10, 11]. In the current
encryption algorithm, which is mainly used for cyber security. It can be divided into
two parts, block encryption [12, 13] and stream encryption [14, 15]. Block cypher
encryption is an encrypt method using the same block cipher key and encrypt one piece
a time to guarantee data security. The block cipher itself can only encrypt a single piece
of data whose length is equal to the length of the cipher block.
The studies of encryption mode [16, 17] have included the integrity protection of
data, that is, the error propagation characteristics of the password in the case where
some data is modified. Subsequent researches [18] identified integrity protection as
another completely different cryptographic goal that was not related to encryption. Part
of the modern working model combines encryption and authentication in an efficient
way, called the authentication encryption mode. Although the working mode is usually
applied to symmetric encryption, it can also be applied to public key encryption.
Stream ciphers (or their principles) have been presented in the world of securing
information for many years. Russian encryption machine called Fialka [19] is an
example of a mechanical ciphering machine used in Cold War. Its design resembles a
stream cipher, because for each letter of plaintext a different encryption, dependent on
previous ones, is used (in other words, two same letters at a different position in a
plaintext are likely to encrypt to two different letters in ciphertext).
Dragon [20] is a new word based stream cipher submitted to the Euro-
pean ECRYPT Stream Cipher Project. It is selected as a Focus Phase 3 candidate as of
April, year 2007. The Dragon is designed with both security and efficiency in mind to
satisfy the need for lightweight algorithms, dedicated to hardware environments where
the available resources are heavily restricted. This stream cipher is constructed on a
1024 bits word-based NLFSR, a state update function (F function) and a 64 bit memory
(M). It is able to generate 64-bit of pseudorandom keystream per round, operating on
key sizes of 128 and 256 bits. The F function plays its role in both key setup and 32-bit
output keystream generation. It is diffused by using a network of modular and binary
addition through G and H function. This design constructed large internal states of
1024-bit and expected period of 2576.
Streaming encryption is a bit level encryption algorithm. Both the encryption and
the decryption use the same pseudo-random stream as the key. Pseudo-random code
[21] means the randomness of the number. However, because the number of cycles is
too large, the size of the sequence of a pseudo-random number column in one cycle is
greater than 1040, so the loop value cannot be seen [22]. Each time the plaintext data is
encrypted correspondingly with the key data stream, a cipher text data stream is
obtained. To encrypt variable length data, the data must first be divided into separate
cipher blocks. In general, the last piece of data also needs to use the appropriate
padding method to extend the data expansion to the length of the cipher block size.
Fialka M-125 [23] is an Enigma based rotor-cipher machine used during the Cold
War. It is considered one of the most secure cipher machines. Advantages of this cipher
are based on the elimination of the Enigma’s known weaknesses. There are no known
attacks on this cipher. The author carried out performance and statistical tests of three
versions of their construction: 4-bit version, 5-bit version and 8-bit version. They
investigated the number of rotors needed to successfully pass the statistical NIST-
testing suite. Their results show that the 4-bit version with 10 rotors and modular
addition/subtraction passes the tests.
In this paper, we present a stream cipher WG-8 [24], which is a lightweight variant
of the well-known WG stream cipher family as submitted to the eSTREAM project.
WG-8 inherits good randomness properties of the WG stream cipher family such as
period, balance, ideal two-level autocorrelation, ideal tuple distribution, and exact
linear complexity. Moreover, WG-8 is able to resist the most common attacks against
stream ciphers including algebraic attack, correlation attack, differential attack, cube
attack, distinguish attack, discrete fourier transform attack, and time-memory-data
tradeoff attack, thereby providing adequate security for lightweight embedded appli-
cations. Lightweight cryptographic primitives are essential for securing pervasive
embedded devices like RFID tags, smart cards, and wireless sensor nodes. WG-8
stream cipher family with 80-bit secret key and 80-bit initial vector (IV), which can be
regarded as a nonlinear filter generator over finite field F28 . More detail will be
described in the next section.
3 Physical and MAC Layer Security
3.1 Received Signal Strength (RSS) of Optical Channel

Received Signal Strength (RSS) approach is the use of the signal from the receiver with
respect to the strength of the relationship between the distance attenuation, with three or
more known fixed position signal source signal strength, after the construction of the
transmission loss patterns in advance, and then after the comparison operation can be
learned from each other the relative distance. In a static optical channel the total
received power Pr at the receiver is given by:
Z
Pr ¼ Pt Hd ð0Þ þ Pt Href ð0Þ ð1Þ
where Hd(0) and Href(0) are the DC channel gain of the direct and reflected paths,
respectively and Pt is total optical power transmitted by LEDs. The DC channel gain of
the direct path is given as:
8 ðm þ 1Þ
< 2pd2 Adet cosm ðhr ÞTs ð#Þgð#Þ cosð#Þ;
Hd ð0Þ ¼ for 0 # #FOV ð2Þ
:
0; for # [ #FOV
where m is the order of the Lambertian radiant, d is the distance between transmitter
and receiver, hr is the angle of irradiance, 0 is the angle of incidence, Ts (0) is the
optical filter gain, g(0) is the optical concentrator gain and #FOV is the FOV of the
receiver. The DC gain of the reflected path can be written as:
8 ðm þ 1Þ
> 2 qAdet dAw cos ðhr Þ cosðaÞ
m
>
< 2ðpd1 d2 Þ
Hd ð0Þ ¼ cosðbÞTs ð#Þgð#Þ cosð#Þ ð3Þ
>
: for 0 # #FOV
>
0; for # [ #FOV
where d1 is the distance between transmitter and reflective point, d2 is the distance
between reflected point and receiver, q is the reflectance coefficient, dAw is a small
reflective area on the wall, a is the angle of incidence from the transmitter and b is the
angle of irradiance from a reflected point. The multipath character of the optical link
can be described by the light error propagation profile.
3.2 Visible Light Transmission Procedure and Duration

In order to secure MAC layer visible light communication, we emphasis is on the
IEEE802.15.7 standard [12], the visible light communication in the media access layer
is based on distributed coordination function (DCF). Therefore, the transmission pro-
cedure will follow the transfer request-to-send (RTS), Clear resend clear-to-send
(CTS), and data recognition data acknowledge (ACK) order. There will be a short inter
frame space (SIFS) between the first packet and the next packet. Therefore, the content
of the optical media access layer will be:
COCC ¼ tRTS þ tCTS þ 3tSIFS ð4Þ
The tRTS and tCTS is the duration time of loss frame while transmission and tSIFS is
specify depletion time of SIFS. Therefore, we can analyze the COCC time and
understand the current packet transmission situation in optical wireless media access
layer once error propagation occurred.
3.3 The Description of the Stream Cipher WG-8

Streaming cypher is a real-time and high accurate encryption algorithm for multimedia
data. However, most of them are missing verification; therefore, the proposed mech-
anism will modify the stream cypher by integrate with pseudo-random code to make
the stream cypher to embedded an authentication function whether the data is modified
during data transmission. As information provided by the EU eSTREAM project, it is
proposed the WG-8 encryption algorithms that conform to have low-power con-
sumption encryption algorithm. These low-power consumption encryption algorithms
included MICKEY, WG-8 [20], Grain, and Trivium. For the best confidentiality, WG-8
has the lowest power consumption compared with rest of three encryption algorithms.
3.3.1 Initialization Phase

The stream cipher WG-8 consists of a 20-stage LFSR with the feedback polynomial l
(x) followed by a WG8 transformation module with decimation d = 19, and operates in
two phases, namely an initialization phase and a running phase. Therefore, the stream
cipher WG-8 is a competitive candidate for securing pervasive embedded applications
(Fig. 1).
Fig. 1. The initialization phase of the stream cipher WG-8
Once the LFSR is loaded with the key and IV, the apparatus runs for 40 clock
cycles. During each clock cycle, the 8-bit internal state S19 passes through the non-
linear WG-8 permutation with decimation d = 19 (i.e., the WGP-8(x19) module) and
the output is used as the feedback to update the internal state of the LFSR. The LFSR
update follows the recursive relation:
Sk þ 20 ¼ ðx Sk Þ Sk þ 1 Sk þ 2 Sk þ 3 Sk þ 4
ð5Þ
Sk þ 7 Sk þ 8 Sk þ 9 WGP-8(S19
k þ 19 Þ; 0 k 40:
After the key/IV initialization phase, the stream cipher WG-8 goes into the running
phase and 1-bit keystream is generated after each clock cycle.
3.3.2 Running Phase

The running phase of the stream cipher WG-8 is illustrated in Fig. 2. During the
running phase, the 8-bit internal state S19 passes through the nonlinear WG-8 trans-
formation with decimation d = 19 (i.e., the WGT-8(x19) module) and the output is the
keystream. Note that the only feedback in the running phase is within the LFSR and the
recursive relation for updating the LFSR is given below:
The WG-8 transformation module WGT-8(x19) comprises of two sub-modules: a
WG-8 permutation module WGP-8(x19) followed by a trace computation module Tr().
Fig. 2. The running phase of the stream cipher WG-8
While the WGP-8(x19) module permutes elements over F82 , the Tr() module com-
presses an 8-bit input to 1-bit keystream.
Sk þ 20 ¼ ðx Sk Þ Sk þ 1 Sk þ 2 Sk þ 3 Sk þ 4
ð6Þ
Sk þ 7 Sk þ 8 Sk þ 9 ; k 40:
1. The keystream has a period of 2160 − 1.

2. The keystream is balanced, i.e., the number of 0’s is only one less than the number
of 1’s in one period of the keystream.
3. The keystream is an ideal two-level autocorrelation sequence.
4. The keystream has an ideal t-tuple (1 t 20) distribution, i.e., every possible
output t-tuple is equally likely to occur in one period of the keystream.
5. The linear span of the keystream can be determined exactly, which is 233.32.
4 Evaluation
In this study, we utilized Raspberry Pi 3B+ board as the experimental test bed. The
Raspberry Pi 3 Model B+ boards have been implemented in many smart factories and
can be represented as a classical embedded system test-bed. In our experiments, a
Raspberry Pi 3+ board was used as the master controller. The master controller is able
to decrypt the data from the cloud server and encrypt the data before sending them to
the cloud server. In our implementations, the experimental results showed that when
both the encryption and data communication functions are enabled, the Raspberry Pi
3B+ CPU was not overrun and could work well. Since the Raspberry Pi 3B+ is not a
high performance computer, this implied that the proposed mechanism could be easily
implemented with low computing power machines. As shown in Fig. 3, the proposed
mechanism will be able to support this low computing power machine and only
increase the total Raspberry Pi 3B+ CPU usage up to 49%. Moreover, once the data
transmission function enabled, the CPU utilization will reached 98%. Even both
encryption and data communication functions are enabled, the CPU utilization will still
running in the reasonable range.
Fig. 3. CPU utilization chart on Raspberry Pi 3 B+
5 Conclusion
In recent years, indoor positioning has draw lots of research attention. Due to the
wireless communication normally have well-known delivery problems, such as signal
fading; multipath propagation, signal obscured and interference problems, and these
will affect the indoor positioning accuracy. Accurate visible light data communication
utilizing received signal strength (RSS) approach that is the use of the signal from the
receiver with respect to the strength of the relationship between the distance attenua-
tion. This approach also can be applied to measurement of optical power absorption
and scattering across an air gap that is to be exposed by harsh factory environment. The
presented paper focus on a model-based understanding of particulates scattering
established, subsequent studies analyzed the impact of visible light scattering on smoke
air and the potential role of dust accumulation. However, future challenges will also
need to be highlight that are associated with visible light scattering measurements of
particulates in the presence of real-world surfaces.
References
1. Zurawski, R.: Industrial communication systems. In: The Industrial Information Technology
Handbook, Sec. 3, pp. 37–47. CRC Press, Boca Raton (2005)
2. Moyne, J.R., Tilbury, D.M.: The emergence of industrial control networks for manufacturing
control diagnostics and safety data. Proc. IEEE 95(1), 29–47 (2007)
3. Benzi, F., Buja, G., Felser, M.: Communication architectures for electrical drives. IEEE
Trans. Ind. Informat. 1(1), 47–53 (2005)
4. HART Field Communication Protocol Specification: HART Communication Foundation
Std., version 7.4 (2007). http://www.hartcomm.org/
5. Vitturi, S., Peretti, L., Seno, L., Zigliotto, M., Zunino, C.: Real-time Ethernet networks for
motion control. Comput. Stand. Interfaces 33(5), 465–476 (2011)
6. Willig, A.: Recent and emerging topics in wireless industrial communications: a selection.
IEEE Trans. Ind. Informat. 4(2), 102–124 (2008)
7. Wireless Systems for Industrial Automation, Process Control and Related Applications,
International Society of Automation (ISA) Standard ISA-100.11a (2009)
8. Sauter, T.: The continuing evolution of integration in factory automation. Proc. IEEE Ind.
Electron. Mag. 1(1), 10–19 (2007)
9. Sauter, T., Lobashov, M.: How to access factory floor information using Internet
technologies and gateways. Proc. IEEE Trans. Ind. Informat. 7(4), 699–712 (2011)
10. Gubbi, J., Buyya, R., Marusic, S., Palaniswami, M.: Internet of Things (IoTs): a vision,
architectural elements, and future directions. Future Gener. Comput. Syst. 29(7), 1645–1660
(2013)
11. Visible Light Communications Consortium. http://www.vlcc.net/
12. IEEE 802.15 WPAN™ Task Group 7 (TG7) Visible Light Communication. http://www.
ieee802.org/15/pub/TG7.html
13. O’Brien, D., Minh, H.L., Zeng, L., Faulkner, G., Lee, K., Jung, D., Oh, Y., Won, E.T.:
Indoor visible light communications: challenges and prospects. In: Proceedings SPIE, vol.
7091 (2008)
14. Pure VLC, Visible Light Communication: An introductory guide (2012). www.purevlc.net
15. Rajagopal, S., Roberts, R.D., Lim, S.-K.: IEEE 802.15.7 visible light communication:
modulation schemes and dimming support. IEEE Commun. Mag. 50(3), 72–82 (2012)
16. Adamko, L., Vojvoda, M., Jokay, M.: Statistical analysis of ECRYPT eSTREAM Phase3
Ciphers, EE časopis pre elektrotechniku a energetiku, pp. 193–196 (2008)
17. Antal, E., Zajac, P.: Analysis of the Fialka M-125 cipher-machine. In: Cryptologia (2013)
18. Engels, D., Fan, X., Gong, G., Hu, H., Smith, E.M.: Hummingbird: ultra-lightweight
cryptography for resource-constrained devices. In: 1st International Workshop on Light-
weight Cryptography for Resource-Constrained Devices. Tenerife, Canary Islands, Spain,
January 2010
19. Rukhin, A., et al.: Sp 800-22 Rev. 1a. a Statistical Test Suite for Random and Pseudorandom
Number Generators for Cryptographic Applications. Technical report, Gaithersburg, MD,
United States (2010)
20. Grosek, O., Vojvoda, M., Zajac, P.: Classical Ciphers. STU (2007). (in Slovak)
21. Perera, T., Hamer, D.: General introduction: Russian cold war era M-125 and M125-3MN
Fialka cipher machines (2005). http://enigmamuseum.com/mfialka.htm
22. Reuvers, P., Simons, M.: Fialka M-125: Detailed Description of the Russian Fialka Cipher
Machines. PAHJ Reuvers & MJH Simons (2009)
23. Robshaw, M., Billet, O. (eds.): New Stream Cipher Designs - The eSTREAM Finalists.
LNCS, vol. 4986. Springer, Heidelberg (2008)
24. Nawaz, Y., Gong, G.: WG: a family of stream ciphers with designed randomness properties.
Inf. Sci. 178(7), 1903–1916 (2008)
Disaster and Emergency Information
Network Systems (IWDENS-2019)
Delay Tolerant Networks with Static Body
Object Detections by Mobile Sensors
for Disaster Information System
Noriki Uchida1(&), Tomoyuki Ishida1, and Yoshitaka Shibata2

1
Fukuoka Institute of Technology, 3-30-1 Wajirohigashi,
Higashi-ku, Fukuoka 811-0214, Japan
{n-uchida,t-ishida}@fit.ac.jp
2
Iwate Prefectural University, 152-52 Sugo, Takizawa, Iwate 020-0693, Japan
Shibata@iwate-pu.ac.jp
Abstract. Although the life safety information services are widely known if
there would be large-scale disasters, there have been some problems to use them
by people who are not good to operate or people who are injured. Therefore, this
paper proposes the static body object detection methods with the DTN for life
safety information system. In the system, the sensors on smartphones auto-
matically detect the abnormal static situations by the time and position differ-
ences based with the Markov Chain Monte Carlo methods, and the emergent
messages are automatically transmitted with the message priorities. Moreover,
those messages are transmitted to the servers by the Data Triage Methods with
the priority ID from the static body detections. Then, the experiments by the city
scale are discussed for the effectiveness of the proposed systems.
1 Introduction
It is now widely known that one can use the life safety information systems by PCs or
smartphones if disasters would occur. For example, the Google Person Finder [1] one
of the well-known services when the one needs the information of the life safety if one
cannot make a contact the relatives after large-scale disasters. Besides, there are cur-
rently many smartphone’s carriers that provide the life information service after the
Great East Japan Earthquake in 2011 [2]. Since the networks would not be available
after large-scale disasters, it is considered that these services are very effective.
However, it is also pointed out some problems of the systems. One of the problems
is that the users need to be good to operate smartphones or PCs. When there is a
disaster, the older people or children tend to be injured according to the previous
earthquakes. Also, if one gets heavily injured by the disaster, the one cannot send the
life safety information by using smartphones or PCs.
Therefore, it is proposed that the Delay Tolerant Networks with Static Body Object
Detections by Mobile Sensors for Disaster Information System [3] in this paper, and
this paper especially reported the simulation results by using the Map of Miyako, Iwate,
Japan. In details, the proposed methods firstly observe the gyro sensors and the
locational data such as GPS or WiFi periodically, and these data are applied for the

https://doi.org/10.1007/978-3-030-15035-8_46
494 N. Uchida et al.
calculation based on the Markov Chain Monte Carlo (MCMC) methods [4] in order to
detect the abnormal static body detections such as injured by disasters. Then, if the
static body is detected, smartphones automatically send the emergent messages by the
DTN (Delay Tolerant Networking) [5] with the Data Triage Methods [6, 7].
In the followings, Sect. 2 describes the network architecture of the assumed life
safety information systems, and Sect. 3 presents the static body object detection
methods and the DTN routing with the priority ID from the static body detections in the
proposed methods. Then, the implementations of the simulator and the experiments are
shown in Sect. 4, and the conclusion and future study are discussed in Sect. 5.
2 Network Architecture
Figure 1 shows the assumed network architecture for the life safety information system
in this research. The network consists of the DTN networks in the damaged areas, the
gateway functions between the DTN and IP networks, and the IP networks in the safety
areas.
Fig. 1. The assumed system architectures of the life safety information system.
If the communication networks would be disconnected or congested by disasters,

the DTN networks establish the message transmissions by the LTE or WiFi D2D
(Device-to-Device) in the damaged areas. Each smartphone or note PC in the areas is
assumed to have the DTN functions, the static body detecting functions, and the
gateway functions to the IP networks. Therefore, if the one gets injured by the disaster,
the emergent messages are automatically created by the detection of the disaster and the
abnormal static body detection, the messages are sent to the other transmittable nodes
with the priority ID by the DTN routing. Then, the messages with the priority ID
reaches to the node that is possible to connect the survived IP network areas, and the
messages are sent to the disaster information servers that are located in the disaster
headquarter or the city hall in the assumed system. On the other hands, if the one is
good to operate the mobile devices and the one is not injured by disasters, the assumed
system can be used as the ordinal life safety information system such as the Google
Person Finder.
Delay Tolerant Networks with Static Body Object Detections 495
3 Proposed Methods
As previously mentioned, it is necessary to consider the victims who were injured or

were not good to operate mobile devices if there would be disasters. Thus, this paper
firstly proposed the static body detection detections [3] for the proposed methods, and
the MCMC methods detect the abnormal static conditions by using the periodically
observed the gyro sensors and the locational data from smartphones. However, it is
supposed that there are two significant problems to detect the static body conditions.
One is the variations of the threshold that is assumed to move in the one’s daily life.
Even if one is staying at the desk or one is sleeping, there are some movements to pick
up a pen or to roll over on the bed. Besides, there are also personal differences between
the daily movements. Secondly, the locational data on the smartphones are different by
the radio environments such as in the basement where the GPS is not available or the
local area where the WiFi is not available. The previous study [3] report the variance of
the locational data on the smartphone in various environments.
Fig. 2. Abnormal static body detections by the time threshold and the movement threshold in
this paper.
Therefore, the static body is detected by the abnormal static state in this research as
shown in Fig. 2. The time threshold and the movement threshold are calculated the
MCMS methods by the periodical observations from the gyro and the locational
sensors on the smartphones, and so these thresholds are optimally changed even if one
moves to the different environments. In the MCMS calculations [4], it approximates the
value function V(s) and the motivation function Q(s, a) by the reward ri as the formula
(1) and formula (2), where a is the learning rate ð0\a\1Þ.
V ðsÞ V ðsÞ þ a½Ri V ðsÞ ð1Þ
Qðs; aÞ Qðs; aÞ þ a½Ri Qðs; aÞ ð2Þ
Here, Ri is also defined as the formula (3).
X
1
Ri ¼ ri þ 1 þ cri þ 2 þ c2 ri þ 3 þ . . . ¼ cr r i þ 1 þ c ð3Þ
r¼0
The MCMS based calculations are held for the setting of the movement threshold
(Rth) and the time threshold (Tth) in the proposed methods for considering the different
environments in the injured victims.
Then, the proposed methods also apply the Data Triage Methods with the priority
ID from the static detections for the DTN routing. The DTN routing [5] is one of the
“store-and-forward” protocols, and so the node stores the received data when there is
no transmittable node around until the transmittable node appeared. However, it is
widely known that there are significant problems such as the large delay or latency, and
so there are various enhanced DTN approaches that are introduced by the Spray and
Wait [8], the Maxprop [9], and the Prophet [10]. The Data Triage Methods for the DTN
routing is also one of the enhanced DTN methods especially considered with the usage
under the disasters. In the Data Triage Methods [6, 7], the messages are once stored in
the queues in the mobile devices as shown in Fig. 3, and the messages are sorted by
order of the proposed priority ID. Here, the priority ID consists of the three digits of the
priority value from the static body detection, the 14 digits of the transmitted data, and
the 64 digits of the Android ID in the transmitted devices. Then, the upper data in the
figure is priority transmitted to the other transmittable nodes by the LTE D2D or the
Android WiFi P2P methods [3].
Fig. 3. Proposed Data Triage Methods with the priority ID from the static body detections
4 Experiments
In the previous paper [3], it was reported that the prototype system is implemented by
Nexus 5X (2 GB MEM, 16 GB Storage, Android OS 6.0.1, IEEE802.11a/b/g/n/ac),
Java 8.0, and Android Studio 2.1.3. Also, the MCMS libraries are customized by the
bases on the DL4j [11] and the HYDRA [12] in the system, and the previous study [3]
indicates the lower latency and the higher delivery rates in the field experiments.
However, the experiments were held on the academic campus, and so this paper
newly evaluated by the simulation with the city map of Taro, Iwate in Japan [6, 7]. In
the evaluations, the map routes of Taro have applied for the ONE (The Opportunistic
Network Environment simulator) ver. 1.6.0 [13], the comparisons between the Spray
and Wait routing and the Maxprop in this paper. Figure 4 shows the implemented map
in the simulator by the proposed methods, and the distance from the place to create
messages to the place to receive the messages is about 5.8 km. In the simulations, the
number of vehicles is differentiated with 25, 50,100, 200, and each vehicle run by a
random walk from 0 km/h to 50 km/h. The radio interface is IEEE802.11b that is the
100 m of the wireless transmission range, and the messages from 0.5 to 1.0 MB are
randomly created by every 50 to 60 s.
Fig. 4. Map image in the simulator by using Taro, Iwate in Japan
The results of the experiments are shown in Figs. 5 and 6. First of all, Fig. 5 present
the results of the delivery rates with the Spray and Wait routing and the Maxprop and
the proposed methods. It is pointed out that the messages with the priority ID in the
proposed methods were completely delivered with 100 vehicles, and that it was the
highest delivery rate than the others. On the other hands, the messages without the
priority ID were lower than the Epidemic and the Maxprop routing. Therefore, it is
supposed that the proposed methods were properly worked to the whole messages, and
that the significant messages were transmitted to the others faster than the non-priority
messages.
Secondly, Fig. 6 shows the results of the latency in the experiments. As the same
with the results of the delivery rates, the priority messages in the proposed methods
were lower latency than the others. On the other hands, the messages without the
priority ID were higher than the Epidemic and the Maxprop routing, but it is supposed
that the proposed methods were properly worked even in these experiments.
Fig. 5. Results of the delivery rate in the simulations by comparison with the Epidemic, the
Spray and Wait, the Maxprop, and the proposed routing methods.
Fig. 6. Results of the latency in the simulations by comparison with the Epidemic, the Spray
and Wait, the Maxprop, and the proposed routing methods.
As the results, it is considered that the proposed methods are effective even by the
city scales, and it might be helpful for the emergent purposes such as the rescues.
5 The Conclusion and Future Study
Although the life safety information services are widely known when there would be
large-scale disasters, it is significant to use them by people who are not good to operate
or people who are injured. Therefore, this paper proposes the static body object
detection methods with the DTN for life safety information system.
In the systems, the static body detections are held by the sensors on the smartphone
with MCMC algorithm, and the messages are transmitted to others by the Data Triage
Methods with the priority ID.
As the results of the simulation by the city scales, it is considered that the proposed
methods are effective for the assumed life safety information systems since the methods
are lower latency and higher delivery rate than these of other DTN routings. For the
future works, the field experiments in the evacuation drills by the prototype system are
planning, and the additional experiments of the static body detections are also planning.
Acknowledgement. This paper is the extend version of “Proposal of Static Body Object
Detection Methods with the DTN Routing for Life Safety Information Systems” in the 32nd
International Conference on Advanced Information Networking and Applications Workshops
(WAINA2018), May. 2018. Also, this work was supported by SCOPE (Strategic Information and
Communications R&D Promotion Programme) Grant Number 181502003 by Ministry of
Internal Affairs and Communications in Japan.
References
1. Google: Google Person Finder. https://www.google.org/personfinder/global/home.html.
Accessed Jan 2018
2. Shibata, Y., Uchida, N., Shiratori, N.: Analysis of and proposal for a disaster information
network from experience of the great East Japan earthquake. IEEE Commun. Mag. 52, 44–
50 (2014)
3. Uchida, N., Shingai, T., Shigetome, T., Ishida, T., Shibata, Y.: Proposal of static body object
detection methods with the DTN routing for life safety information systems. In: The 32nd
International Conference on Advanced Information Networking and Applications Work-
shops (WAINA2018), pp. 112–117 (2018)
4. Andrieu, C., Freitas, N.D., Doucet, A., Jordan, M.I.: An Introduction to MCMC for Machine
Learning. Kluwer Academic Publishers, Boston (2003)
5. Fall, K., Hooke, A., Torgerson, L., Cerf, V., Durst, B., Scott, K.: Delay-tolerant networking:
an approach to interplanetary internet. IEEE Commun. Mag. 41(6), 128–136 (2003)
6. Uchida, N., Kawamura, N., Shibata, Y., Shiratori, N.: Proposal of data triage methods for
disaster information network system based on delay tolerant networking. In: The 7th
International Conference on Broadband and Wireless Computing, Communication and
Applications (BWCCA2013), pp. 15–21 (2013)
7. Uchida, N., Kawamura, N., Sato, G., Shibata, Y.: Delay tolerant networking with data triage
method based on emergent user policies for disaster information networks. Mobile Inf. Syst.
10(4), 347–359 (2014)
8. Spyropoulos, T., Psounis, K., Raghavendra, C.S.: Spray and wait: an efficient routing scheme
for intermittently connected mobile networks. In: Proceedings of the 2005 ACM SIGCOMM
Workshop on Delay-Tolerant Networking, WDTN 2005, pp. 252–259 (2005)
9. Burgess, J., Gallagher, B., Jensen, D., Levine, B.: MaxProp: routing for vehicle-based
disruption-tolerant networks. In: Proceedings of the 25th IEEE International Conference on
Computer Communications, INFOCOM 2006, pp. 1–11 (2006)
10. Lindgren, A., Doria, A., Scheln, O.: Probabilistic routing in intermittently connected
networks. ACM SIGMOBILE Mobile Comput. Commun. Rev. 7(3), 19–20 (2003)
11. DL4 J: Markov Chain Monte Carlo and Machine Learning - Deeplearning4j. https://
deeplearning4j.org/markovchainmontecarlo. Accessed Feb 2018
12. Warnes, G.R.: HYDRA: a Java library for Markov Chain Monte Carlo. J. Stat. Softw. 7(4)
(2002).
13. The ONE. https://akeranen.github.io/the-one/. Accessed Jan 2019
A Basic Study on Emergency Communication
System for Disaster Using LPWA
Hiroaki Yuze1(&) and Shinichi Nabeta2

1
University of Shizuoka, Shizuoka, Japan
yuze@u-shizuoka-ken.ac.jp
2
Johnan Shizuoka High School, Shizuoka, Japan
s-nabeta@fw.ipsj.or.jp
Abstract. When disasters occur, there is a high possibility of road collapse,

communication disruption, blackouts in the mountain area compared to the flat
ground. We considered the use of LPWA as an emergency communication
means in mountain areas. Mountains are a shield for radio waves, but we plan to
expand communication range by installing a relay server of LPWA near the
summit. In this study, we conducted unidirectional communication experiment
using LPWA in mountain area in northern part of Hamamatsu city and con-
firmed the communication range of LPWA. Also, we install a relay server of
LPWA in the middle of the mountain and confirmed that message transmission
is possible even to the extent that we cannot see it. For long distance commu-
nication for relay servers, we install a relay server at University of Shizuoka
conducted a communication experiment of LPWA, confirmed that we can
communicate 38 km ahead.
1 Introduction
There are many natural disasters in Japan, and disasters such as the Osaka prefecture
northern quake, the western Japan torrential rainfall, and the Hokkaido Ballistic Eastern
Earthquake occurred even in 2018 alone. When large-scale disasters occur, there is a
possibility that traffic is blocked due to damage to road structures, or communication
infrastructure is damaged, and communication is interrupted. Moreover, there is a high
possibility that power outages will last for a long time due to the damage of trans-
mission lines and others. Especially in the mountain area, the road is cut off due to cliff
collapse, etc., compared to the plain area, the isolation of the village, the possibility of
communication disruption and power outage is high. A study [1] is being conducted to
utilize amateur radio as a measure against isolation between mountainous areas. In
recent years, LPWA has attracted attention as a technology capable of communicating
over a wide range with low power consumption. In this research, we consider LPWA to
be used as a means of emergency communication when the communication infras-
tructure in the mountain area cannot be used due to disaster. In this research,
we propose to use LPWA which is more restrictive than the usual Internet commu-
nication in mountainous areas. We examined a method to utilize it for emergency
communication.

https://doi.org/10.1007/978-3-030-15035-8_47
502 H. Yuze and S. Nabeta
We conducted a communication experiment conducted to investigate the radio

wave characteristics of LPWA in the mountain area. We also installed a relay server of
LPWA in the middle of the mountain, and also conducted experiments on message
transmission via a relay server. Furthermore, in order to examine the possibility of
long-distance communication for relay servers, we conducted an experiment of long-
distance communication using LPWA and a relay server. In this paper, we will discuss
the discussion on utilization of LPWA as an emergency communication means in
mountainous areas. We will also describe the results of communication experiments
using LPWA.
2 The Utilization of LPWA in Emergency Communication
LPWA is an acronym for Low Power Wide Area. There are Bluetooth, ZigBee and so
on such as the art of low-power communication. LPWA can run more low-power those
arts and provides long-range communication. LPWA is ultra-narrow band, but LPWA
attracts a remark for IoT/M2M device, which is expected becoming popular, con-
necting with network. The main installation specifications of LPWA include LoRa,
SIGFOX, NB-IoT, and so on. In Japan, the usage with LoRaWAN method which is
used LoRa increases [2].
Generally, we can think that it uses LPWA as the means of collecting each sensor
information as the utilization for disaster prevention of LPWA. There is an experiment
of the usage in LPWA in the acquisition of the data of the level-gage of the underline
and the checker of the exposure-to-water by the inundation inside a levee in the
canalization. There is an experiment of the outgoing of the information on the disaster
prevention administration wireless which utilized LPWA [3].
In this study, we think of utilizing LPWA for an emergency communication in the
isolated mountainous area. We need a license for usage of the amateur radio, but
LPWA can be used by without a license. When using as means of communication in
self-, mutual help in the disaster, LPWA become means of communication which
anyone can use because the license is unnecessary.
3 Considering for LPWA Communication in Mountainous

Area
The mountainous area is different from the plains and doesn’t often put on by the
direct-communication, foreseeing the end of the communication because there are
mountains as blocker. But there is possibility that the mountain reflection or mountain
diffraction can be used for the radio-propagation. When installing an antenna in the
heights in the plains, it is necessary to build the steel tower and so on. But when
installing an antenna in a summit or a hillside in mountainous area, we can wireless
communication in the heights.
In spite of short distance communication, it is difficult to communicate with direct-
communication in Low-Low place wireless communication like Terminal A-Terminal
B in Fig. 1. Because of installing a wireless repeater in a summit or a hillside, we can
A Basic Study on Emergency Communication System for Disaster Using LPWA 503
communicate directly without foreseeing like Terminal A-Repeater-Terminal B com-

munication in Fig. 1. Installing repeater in heights, repeater cannot only repeat com-
municate, but also use mountain reflection.
The band of LPWA is too ultra-narrow to send long data. When we communicate
for a disaster, it is necessary to use short message like a code which is preparing for
emergency communication. And we can send some message only selecting a code.
When we think of sending a message without a code, we slice a message.
Repeater
Terminal A Terminal B Terminal C
Fig. 1. The image of communication using LPWA in mountainous area
Using LPWA communication in mountainous area has the advantage of much

smaller size because of using smaller power-unit like a battery or solar panel. Then, we
can install or carry with ease.
4 Experiment Environments for LPWA Communication
We use “RM-EVSA-92AS-2” which is a simple test tool for LoRa/FSK distribution

from Green House [4] in an experiment for LPWA communication. The specifications
given in Table 1 and appearance given in Fig. 2. We use our own development pro-
grams for this experiment.
Table 1. Specifications for RM-EVSA-92AS-2 [4]

RF specs Frequency 920 MHz
Max Tx power +13 dbm (Approx. 20 MW)
Sensitivity −137 dbm
Supported RF Proprietary/IEEE802.15.4g compliant
Modulation LoRa/FSK/GFSK (GFSK under development)
Power consumption Tx: 28 mA, Rx: 10.5 mA, Standby: below 1 lA
Transfer speed 292.97–300 Kbps
External Weight Approx. 60 g
specs External dimensions W 82.0 D 22.0 H 65.0 (mm) (excluding the
(Case) protuberance and antenna)
Fig. 2. Appearance of RM-EVSA-92AS-2 [4]
Our own development programs are divided into three programs, for sending,
repeating, and receiving. And for sending program is divided in two programs for
sending free message and for sending auto increment number for communication test.
For repeating program sends the just receiving message. For repeating and sending
programs have sender information in header on the message. And we can see the
message and judge from whether repeater or terminal using for receiving program. In
addition, we development reading receive message function by using Open JTalk [5]
for using in driving a car. These programs are output logs (CSV) about each
send/receive message and time, RSSI and so on.
5 LPWA Communication Experiment 1 in Mountainous

Area
We experimented for LPWA communication in January 3rd and 6th–7th, 2018 at

northern part of Hamamatsu city, JAPAN. The purpose of this experiment is to confirm
whether communication is possible between places which cannot be directory com-
municated directly by installing an LPWA communication device having a relay
function in a place with heights and high prospects.
The location where the communication device was actually installed is shown in
Fig. 3 (left). The map shown in Fig. 3 (left) is a marker of each site added to the
electronic map (tile) from Geospatial Information Authority of Japan. Markers in the
figure, “●” indicates Akiba Dam, “○” indicates Aoya Pocket Park, “” indicates
Hamamatsu City Tatsuyama Forest Culture Center, and “★” indicates Shimohirayama
Noson park. And the state of the sending side (Akiba Dam) at the time of communi-
cation experiment is shown in Fig. 3 (right).
Fig. 3. The location of LPWA communication experiment in mountainous area (left), and state
of sending side (Akiba Dam)
On January 3rd, we confirmed the location where communication is

possible/impossible based on Akiba Dam. As a result, when we sent data to Aoya
Pocket Park on the receiving side five times from Akiba Dam on the sending side, we
received all the data. The average value of RSSI was −134, and it was confirmed that
communication is possible between Akiba Dam and Aoya Pocket Park. On the other
hand, it was found that Akiba Dam and Hamamatsu City Tatsuyama Forest and Culture
Center could not communicate. Both points are where we cannot see the Akiba Dam
directly, but we could communicate at Aoya Pocket Park because the distance was
short.
Fig. 4. State of sending side (Shimohirayama Noson Park) (left), and state of receiving side
(car) (right)
On January 6th to 7th, we have grasped the range of communication from high
altitude prospects. At that time, based on Shimohirayama Noson Park, we installed
LPWA device in the car and confirmed communicability while driving car. Figure 4
(left) shows a state of Shimohirayama Noson Park as a sending side. Figure 4 (right)
shows a state of car as a receiving side.
Figure 5 shows the locus of movement of the car moving from Shimohirayama
Noson Park for reception. It is the point where the red dot in the figure was recorded as
the GPS logs. In Fig. 5 the big dots were receivable places, the small dots were not. As
a result of receiving experiments while driving by car, it was confirmed that both Akiba
Dam and Hamamatsu Tatsuyama Forest and Culture Center can communicate from
Shimohirayama Noson Park.
Shimohirayama Noson Park
Fig. 5. Sending experiment GPS log from Shimohirayama Noson Park
6 LPWA Communication Experiment 2 in Mountainous

Area
On February 18th, 2018, we conducted an LPWA communication experiment in the

northern part of Hamamatsu city. We set up an LPWA device at Fudo Falls and
confirmed while driving by car with how far we can communicate. After that, we
moved to Shimohirayama Noson Park, established an LPWA relay station there and
moved to Akiba Dam. We conducted a relay experiment of message communication
with Akiba Dam as the sending point of the message. We installed the LPWA device as
the receiving side in the car and carried out a relay experiment of message commu-
nication while driving car. We were able to input and send some messages from Akiba
Dam with PC, send it again at the relay device at Shimohirayama Noson Park, and
receive the relayed message at the Hamamatsu Tatsuyama Forest and Culture Center.
The received message is displayed on the PC on the receiving side, and the PC
synthesizes the message and reads it. By sending several patterns of messages, we
confirmed that we can send messages without problems between Akiba Dam and
Hamamatsu Tatsuyama Forest and Culture Center. After that, we further changed the
reception location and confirmed how far the message communication can be done.
Figure 6 shows the message displayed on the receiving computer.
Fig. 6. State of receiving message
7 Long Distance Communication Experiment Using LPWA
University of Shizuoka to which the authors belong is located at the foot of Nihondaira,
the altitude is about 100 m, and it is located in a well-prospective place. It is also the
self evacuation area and fully autonomous anti-disaster base station [6, 7] have also
been established. We conducted a communication experiment between university and
Shizuoka prefecture disaster hospital with a wireless network. Therefore, we conducted
a communication experiment on the actual communication range of the LPWA device
on September 17th, 2018.
In the University of Shizuoka as the sending side, we set up a PC that launched the
LPWA device and the sending program which is only sent auto increment number and
in the car as the receiving side, we installed the LPWA device. We carried out reception
experiment while moving by car. As a result, the maximum coverage area in this
experiment was near the Surugawan-Numazu Smart Interchange on the Shin Tomei
Expressway, which is a straight distance of 37 km from University of Shizuoka. This
result is shown in Fig. 7. The map shown in Fig. 7 is a marker of each site added to the
electronic map (tile) from Geospatial Information Authority of Japan. In Fig. 7, “●”
indicates University of Shizuoka and “★” indicates the maximum communicable point
in this experiment.
Fig. 7. Maximum communicable point
8 Conclusion
In this research, we studied a method to utilize LPWA as an emergency communication

method at the time of a disaster, actually conducted a communication experiment at the
mountain area, and grasped the LPWA communication characteristics at the mountain
area. We set up a relay server of LPWA in the middle of the mountain and confirmed
the effectiveness of the relay server in the mountain area. Also, we developed a pro-
totype system of emergency communication using LPWA with very low data com-
munication speed and used it for communication test in the mountain area, confirmed
that message transmission and table and its transmission voice synthesized output can
be done. In order to increase the usefulness of emergency communication using the
relay server of LPWA, we conducted long distance communication experiment of
LPWA considering long distance communication toward relay server. We set up a
relay server of LPWA at the university in the hill and confirmed that we can com-
municate from about 37 km ahead.
From now on, I would like to aim for practical application of emergency com-
munication using LPWA by making use of the communication experiment result of
LPWA obtained in this research and the evaluation result of prototype system.
References
1. Ueno, K., Mori, A., Nakano, S., Yoshida, A.: Utilization of amateur radio in isolated
mountainous village due to disaster. In: Proceedings of the JSCE Earthquake Engineering
Symposium, vol. 30 (2009). 4-0036
2. Tanaka, M.: Recent development of LoRaWAN in Japan, IEICE Technical report, vol. 117,
no. 297, SRW2017-58, pp. 59–64 (2017)
3. Yuze, H., Shibata, Y., Sugawara, K., Nakagawa, Y., Ono, M.: Communication experiment on
information dissemination by disaster prevention administrative radio with wireless
communication technology “LPWA”. In: Proceedings of the 19th Conference, Japan Society
for Information Studies, pp. 120–121 (2017)
4. Green House: Development board, RM-EVSA-92AS-2. https://www.green-house.co.jp/
products/rm-evsa-92as-2/
5. Open JTalk: Open Jtalk. http://open-jtalk.sourceforge.net/
6. Yuze, H., Nabeta, S., Ito, Y.: Development of autonomous anti-disaster base station by
combination of satellite and wireless networks. In: 2015 Workshops of 29th International
Conference on Advanced Information Networking and Applications (WAINA), pp. 524–528
(2015)
7. Yuze, H., Nabeta, S., Ito, Y., Kosaka, H., Shibata, Y.: Improvement of autonomous anti-
disaster base station by combination of satellite and wireless networks. In: 2016 Workshops of
30th International Conference on Advanced Information Networking and Applications
(WAINA), pp. 667–672 (2016)
Platform System Based on LoRa Mesh
Network Technology
Goshi Sato1(&), Yoshitaka Shibata2, and Noriki Uchida3

1
National Institute of Information and Communications Technology,
4-2-1 Norikibe-cho, Koganei, Tokyo 184-8795, Japan
sato_g@nict.go.jp
2
Faculty of Software and Information Science, Iwate Prefectural University,
152-52 Sugo, Takizawa, Iwate 020-0693, Japan
shibata@iwate-pu.ac.jp
3
Higashi-Ku, Fukuoka 811–0295, Japan
n-uchida@fit.ac.jp
Abstract. Although the LoRa private network has started to be used in various
applications due to its characteristics, there are only arrangements for carrier
sense time, transmittable time, and post-transmission standby time on the LoRa
standard and Radio Law, It is necessary for the user side to independently deal
with behaviors related to retransmission of data lost at the time of collision and
error correction function. In addition, since it is a relatively new wireless
communication method, it can be said that it is almost in a disorderly state
without unified guidelines and tacit agreements that have been accumulated in
the usual way. This situation is particularly problematic when various services
start utilizing LoRa with their own protocols, so a uniformly available LoRa
platform is necessary.
1 Introduction and Purpose
The purpose of this research is to construct a unified LoRa self-employed network

platform that can be used in various situations using LoRa mesh network technology.
Specifically, by applying the LoRa mesh network technology to design the optimum
setting and data structure in the following three communication environments, it is
possible to realize communication in which the example application is sufficiently
operable in various communication environments with different situations It is aimed to
evaluate that performance can be demonstrated through experiments.
1. Mobile phone network Sharing road conditions by inter-vehicle/road-vehicle
communication on roads crossing insensitive areas
2. Temporary observation point data transfer and collection in mountainous areas
Such a platform that meets various communication environments and application
requirements is one that could not be realized in the LoRa private network system so
far and is realized by utilizing the LoRa mesh network technology of our organization.

https://doi.org/10.1007/978-3-030-15035-8_48
Platform System Based on LoRa Mesh Network Technology 511
2 Methods
In order to exchange and share information with neighboring shoulder system in real
time, we will integrate N different wavelengths of different wavelengths and develop N
wavelength cognitive radio. Then, the radio link that maximizes the evaluation function
based on the necessary communication quality (throughput, delay, packet loss rate,
RSSI) is determined according to the change of the communication environment
(communication distance, noise ratio, interference, etc.). By maximizing communica-
tion distance and total data transfer amount by realizing inter-vehicle communication.
For this reason, we develop a function that can select an optimal wireless link from N
wavelength radio links by SDN (Software Defined Network) protocol function. As the
N wavelength radio, a radio for IoT such as a wireless LAN or LoRa radio corre-
sponding to the IEEE802.11n (2.4 GHz)/ac (5.6 GHz)/ad/60 GHz) standard that is
available for license generation and is the next generation wireless network We develop
an N wavelength cognitive radio unit by suitable combination from standards etc. In
particular, LoRa wireless communication uses broadcast communication, if collision of
wireless signals occurs due to simultaneous transmission, retransmission is not per-
formed. Also, in consideration of various communication environments, nodes that
perform communication have hidden terminal problems because radio wave reach
ranges are extremely different. In order to avoid simultaneous transmission, LoRa mesh
network technology implements time-division access control method after time syn-
chronization by GPS. It is assumed that all the radio stations operated in the system are
numbered by ID and the total number is also grasped. In this premise, based on the ID
dynamically allocate time slots that it can transmit. For example, assuming that the
transmission time per wireless device is 10 s (considering the time required for backoff
to avoid collisions with other systems), when using five wireless stations, these five
devices do not have any problems A time longer than 50 s that can be transmitted is
assigned as a transmission cycle. In this transmission cycle, as shown in Fig. 1,
transmission time slots are assigned based on respective IDs. As a result, occurrence of
collision in communication in the own system is avoided. The amount of data that can
be transmitted per unit time of the wireless device and the communicable distance vary
depending on factors such as the spreading factor of LoRa, the used channel band-
width, the surrounding communication environment, and the like. Also, since the
transmission cycle varies according to the number of radio devices used in the system,
slot design of the entire system is required in advance according to the requirements.
Therefore, in this research, by appropriately designing the spreading factor of the
LoRa, the used channel bandwidth, and the transmission cycle according to the fol-
lowing three kinds of application systems on the basis of the LoRa mesh network
technology, Verify feasibility. By the way, data relaying is done by using flooding.
However, in addition to the mechanism to avoid transmitting the same information
multiple times, in order to reduce the number of flooding transmission, the following
implementation is carried out.
First, by sending information listing the sender IDs of the signals received by each
node in the past as control signals at regular intervals, it is possible to grasp the range
within which their neighboring nodes can communicate. When a newly received
512 G. Sato et al.
Fig. 1. Relationship between transmission cycle time and transmission time per wireless device
information is received, if there is a node that advertised the information earlier before
its own transmission timeslot comes around, if the neighboring node that it covers is
adjacent to If it is all covered, it will not retransmit. As a result, retransmission of
unnecessary information is suppressed.
2.1 Road Situation Sharing by Road-to-Vehicle Communication

on a Road Crossing a Mobile Phone Network Insensitive Area
The proposer has been researching the speeding up of WiFi communication at the time
of mistake by pre-sharing of WiFi connection information using LoRa. In order to
further develop this, we will develop an in-vehicle type road-vehicle communication
system using LoRa mesh network technology that enables bidirectional data storage
type communication by LoRa. In advance, share meta information such as the type of
application information held by each vehicle, the type of application data requested, the
direction of travel of the vehicle, the traveling history, etc., and automatically determine
the information to be transmitted, It is possible to more efficiently allocate the com-
munication time by WiFi at the time of limited mismatch to data transfer. Unlike
communication between fixed stations, communication time at mismatch is limited also
in LoRa communication in road-to-vehicle communication, so design the transmission
cycle time as short as possible as shown in Fig. 2 below So, share meta information
efficiently.
Fig. 2. Transmission cycle time design for inter-vehicle/road-vehicle communication

In this research, in order to limit the maximum number of terminals simultaneously

communicating, the target vehicle is assumed to be a called route bus, and the loadside
server is installed at the bus stop or the like. Evaluate the amount of data transfer at the
time of passing.
2.2 Transfer and Collect Data Between Temporary Observation Points

in the Mountain Area
We develop an IoT sensor data collection network system using LoRa mesh network
technology as a system for instantly constructing an information communication net-
work enabling planar and real-time measurement by sensors in a limited mountainous
area. Among the collected sensor data, there is a relatively large data amount. In order
to cope with such a difference in size of data to be collected, it is possible to variably
secure one slot time on the IoT sensor data collection network system side as a
transmission path as shown in Fig. 3.
Fig. 3. Variable slot overview for collecting sensor data with different data amount
2.3 Link Selection Method

To the problem of degraded network performance due to congestion and so on, in this
research, we solve by link switching based on network performance. In the proposed
method, to detect the performance degradation of the network by monitoring on a
throughput, delay, and packet loss. First, the measurement results are smoothed by this
formula.
ni ¼ amt þ bmt1 þ cmt2
a, b, c are the weight of moving average. m is the measurement results for each
time.
Then, smoothing values are normalized by this formula.
8 ni li
>
< ð1 ui li Þ 10 ðli \ni \ui Þ
Si ¼ 1 ðni ui Þ
>
:
9 ðni li Þ
ni l i
Si ¼ ð Þ 10
umax li
514 G. Sato et al.
n is the smoothed value, l is the minimum value, and u is the maximum value in
monitoring history. Finally, it will be evaluated using the normalized value by this
formula.
Pi ¼ xnthroughput þ ynpacketlossrate þ zndelay
x, y, z are the weight of the network policy. Basically, use the link with the best Pi
value. However, since there is a characteristic that the throughput is very low with
respect to the LoRa radio, basically it is used like a control channel for exchanging
information for the above judgment and key information for connection.
3 Preparation of Experimental Environment
For field demonstration experiment from next fiscal year, we are currently studying
outdoor prototype system as below (Fig. 4).
Fig. 4. Prototype system overview
In addition, prototype N-wavelength wireless communication prototype system

(Figs. 5 and 6) implementing this protocol is prototyped using NerveNet [1] developed
by NICT and a laboratory test is being conducted. Basic protocol operation can be
confirmed by expected operation, basic data collection and system design of N
wavelength cognitive radio, which is the target of this fiscal year, system design,
implementation per module are expected to be completed in the current fiscal year.
Implementation to combine IEEE802.11 bgn and LoRa will be implemented in the
future, and information transfer experiments using various data and roadside-vehicle
communication will be demonstrated outdoors.
Fig. 5. System appearance (on Fig. 6. System appearance

vehicle) (loadside)
4 Summary
In this paper, we propose a unified LoRa platform that does not rely on existing
communication infrastructure, and designed to adapt to time-division access control
method necessary for realizing it. In addition, we developed a system for sharing
vehicle position information using a self-operated network using a basic time division
access control method, actually installed a radio at the hospital rooftop of the city
center, and then sent out-of-line communication We verified whether or not desired
communication can be done even in environmentalization. As a result, it was confirmed
that even in the downtown area, it is sufficiently possible to communicate about the
sharing of the position information of the vehicle within a radius of about 2 km. In
addition, it is shown that time sharing access control communication using basic
technology of LoRa platform enables simultaneous sharing of information among
multiple sites.
Acknowledgments. The research was supported by Strategic Information and Communications

R&D Promotion Program Grant Number 18152003 by Ministry of Affairs and Communication
in Japan.
Reference
1. Inoue, M., Ohnishi, M., Peng, C., Li, R., Owada, Y.: NerveNet: a regional platform network
for context-aware services with sensors and actuators. IEICE Trans. Commun. E94-B(3),
618–629 (2011)
Construction of a Disaster Response Automatic
Extraction Support System
Tatsuya Ohyanagi1, Tomoyuki Ishida2(&), Noriki Uchida2,

and Yoshitaka Shibata3
1
Ibaraki University, Hitachi, Ibaraki 316-8511, Japan
17nm704r@vc.ibaraki.ac.jp
2
Fukuoka Institute of Technology, Fukuoka, Fukuoka 811-0295, Japan
{t-ishida,n-uchida}@fit.ac.jp
3
Iwate Prefectural University, Takizawa, Iwate 020-0693, Japan
Abstract. Japan is a country of frequent natural disasters, and large-scale

natural disasters occurs every year. Therefore, we conducted a hearing survey of
disaster response to local governments and summarized the disaster response
process. As a result, we found that the municipal disaster countermeasures
headquarters is spending a great deal of time collecting and organizing various
disaster information. Therefore, in order to solve the issues of municipal disaster
countermeasures headquarters, we have constructed the disaster response
automatic extraction support system. The disaster response automatic extraction
support system accumulates various disaster information such as disaster
response record of local governments and local disaster prevention plan. In
addition, this system supports decision-making by extracting past disaster
response records according to the current situation of municipal disaster coun-
termeasures headquarters. This system consists of a disaster information accu-
mulating system for accumulating data and a disaster information visualization
and analysis system for visualizing and analyzing accumulated disaster infor-
mation. The disaster information visualization and analysis system also visu-
alizes and analyzes social media information.
1 System Configuration
In order to solve the issues of previous research [1, 2], we proposal the disaster
response automatic extraction support system. The configuration of the system con-
structed in this research is shown in Fig. 1. This system supports decision-making of
municipal disaster countermeasures headquarters installed in each local government at
the occurrence of a disaster. In normal times, the user accumulates past disaster
response records and local disaster prevention plans in this system. On the other hand,
in emergency times, the user searches for disaster response records and shares infor-
mation by using the accumulated data. Also, the user visualizes registered disaster
information and social media information related to the disaster. In this research, we
refer to the “disaster information accumulating system” for the system used at normal
times and the “disaster information visualization and analysis system” for the system

https://doi.org/10.1007/978-3-030-15035-8_49
Construction of a Disaster Response Automatic Extraction 517
used at emergency times. The disaster information accumulating system is used by

disaster prevention staff of each local government. And, the disaster information
visualization and analysis system is used by staff who organize and analyze disaster
information.
Fig. 1. The system configuration in the disaster response automatic extraction support system.
518 T. Ohyanagi et al.
The disaster information accumulating system and the disaster information visu-
alization and analysis system communicate with the Disaster Historical Records
Visualization and Analysis System API (DHRVAS API) on the server for data man-
agement and visualization. By communicating with the DHRVAS API, this system
realizes reduction of the amount of data received. In addition, this system realizes
improvement of the response speed of the system and reduction of communication
load. In order to realize this function, we use asynchronous communication technology
and standard data format. We adopted Fetch API provided by JavaScript in asyn-
chronous communication technology and JSON (JavaScript Object Notation) in stan-
dard data format. In addition, we have increased system extensibility and diversion by
SOA (Service Oriented Architecture) which clearly divides server application and
client application. For example, when adding a new function, it is possible to add the
logic to the server and use the function by simply calling the API on the client side.
Moreover, by calling the API, it is easy to facilitate data linkage with other systems.
This also facilitates the introduction of new systems. The method that the existing
system uses the data of this system via the API is shown in Fig. 2.
Fig. 2. API cooperation between this system and other systems
This system uses Relational Database (RDB) and NoSQL Database (NoSQL DB)
as database server. The RDB consists of disaster classification table, disaster prevention
plan table, multimedia table, disaster table, disaster response record table, disaster
response phase table, region table, prefecture table, municipal table, user table, and
disaster category table. The ER diagram of each table is shown in Fig. 3.
Fig. 3. ER diagram of RDB in this system
2 Full-Text Retrieving Engine
In this research, various kinds of disaster information such as disaster name, disaster
response record, local disaster prevention plan etc. are accumulated in the database.
This will promptly provide past disaster information to users in emergency times and
support disaster response decision making. In order for the user to extract past disaster
information, data is retrieved by using parameters such as date and time, disaster
classification, disaster response phase, and the like. Furthermore, this system provides a
full-text retrieving function that extracts texts most approximate to the input search
keywords. Full-text retrieve is one of the document retrieval methods and is different
from general retrieval methods (progressive scan retrieve and sequential retrieve) for
retrieving documents one by one. By scanning all documents in advance and con-
structing index data, high-speed retrieve is realized. The outline of full-text retrieve is
shown in Fig. 4.
Fig. 4. The outline of full-text retrieve

In order to realize full-text retrieve, it is necessary to construct an index used for

full-text retrieve. In incorporating this full-text retrieving function into our system, we
formatted the data before doing the indexing process. The procedure of the processing
necessary for realizing full-text retrieve by formatting data is as follows.
(1) Standardization of text information
(2) Split into morphemes
(3) Unification of character types
(4) Numeric substitution
(5) Stop word removal
3 Disaster Response Automatic Extraction Support System
The disaster response automatic extraction support system consists of the disaster
information accumulating system and the disaster information visualization and anal-
ysis system.
3.1 Common Functions of Disaster Response Automatic Extraction

Support System
In using this system, the user must first perform user authentication (login). In this
research, we realized the authentication function using JWT authentication. A login
screen for user authentication is shown in Fig. 5. On the login screen, the user inputs a
user ID and a password. Thereafter, by pressing the “Login” button, authentication
processing is performed. If the information inputted by the user is valid, the user
authentication is completed and the screen transitions from the login screen.
Fig. 5. User authentication (login) screen
Next, the logout function is displayed by selecting the menu button installed at the
top of the main screen of the system with the user logged in (Fig. 6). When the logout
button is pressed, the user information and all the information cached by the system are
deleted.
Fig. 6. Logout screen of the disaster information accumulating system and the disaster
information visualization and analysis system
3.2 Disaster Information Accumulating System

After completion of the authentication process, the login screen transits to the basic
information selection screen. The basic information selection screen of the disaster
information accumulating system is shown in Fig. 7. On this screen, the user selects the
basic information necessary for using the disaster information accumulating system.
A form is provided for entering “disaster type” and “local government name” in the
screen, and the user inputs information into these forms. The user can select disaster
type of earthquake disaster, tsunami disaster, wind and flood damage, volcanic disaster,
snow damage. Since this system realizes use of location information, the local gov-
ernment name based on the location information is automatically entered. After
completing the data input, pressing the confirm button transits to the next screen. At
this time, the information entered on this screen and the user’s authentication infor-
mation are stored in the local storage of the Web browser.
Fig. 7. Basic information selection screen of the disaster information accumulating system
3.3 Disaster Information Visualization and Analysis System

Like the disaster information accumulating system, the disaster information visual-
ization and analysis system has common data to be used by all functions in the
visualization and analysis system. The basic information selection screen of the disaster
information visualization and analysis system is shown in Fig. 8. On the basic infor-
mation selection screen of the disaster information visualization and analysis system,
the user selects “disaster response phase” and “season” in addition to “disaster type”
and “local government name” which were installed in the disaster information accu-
mulating system. The season is automatically calculated from the current date and time
by the system. After completing the data input, pressing the confirm button transits to
the next screen. At this time, like the disaster information accumulating system, the
information entered on this screen and the user’s authentication information are stored
in the local storage of the Web browser.
Fig. 8. Basic information selection screen of the disaster information visualization and analysis
system
4 Conclusion
In this research, we conducted a hearing survey on local governments and clarified the
current disaster response process. As a result, we found that the municipal disaster
countermeasures headquarters is spending a great deal of time collecting and orga-
nizing damage information and shelter information. This hinders prompt decision
making within the municipal disaster countermeasures headquarters. Therefore, in this
research, we extracted requirements for smoothly making decisions from the disaster
response process obtained in the hearing survey. Based on this requirement, we have
constructed the disaster response automatic extraction support system aiming to sup-
port decision making of municipal disaster countermeasures headquarters.
Acknowledgments. This research was supported by JSPS KAKENHI Grant Number

JP16K00119.
References
1. Takahagi, K., Ishida, T., Sakuraba, A., Sugita, K., Uchida, N., Shibata, Y.: Construction of a
mega disaster crisis management system. J. Internet Serv. Inf. Secur. (JISIS) 5(4), 20–40
(2015)
2. Ishida, T., Hirohara, Y., Uchida, N., Shibata, Y.: Implementation of an integrated disaster
information cloud system for disaster control. J. Internet Serv. Inf. Secur. (JISIS) 7(4), 1–20
(2017)
Network Performance Evaluation to Realize
N-Wavelength V2X Cognitive Wireless
Communication System
Akira Sakuraba1(&), Takanori Ito2, and Yoshitaka Shibata1

1
Regional Cooperative Research Center, Iwate Prefectural University,
Takizawa, Iwate Prefecture, Japan
{a_saku,shibata}@iwate-pu.ac.jp
2
Faculty of Science and Engineering, Iwate University,
Morioka, Iwate Prefecture, Japan
takaito@iwate-u.ac.jp
Abstract. Intelligent Transport Systems realize safer driving which is consid-

ered by road state information from multiple types of onboard or roadside
sensors. Autonomous driving vehicle essentially requires to collect and process
own sensor output and operate autonomously. When onboard system processes
sensor output for onboard decision of road conditions at the position, it is
necessary to consider how the vehicle exchange road information with other
vehicle or roadside unit. This paper introduces cognitive wireless communica-
tion method which is designed for exchanging of road sensing information. Our
proposal system consists of 920 MHz, 2.4 GHz and 5.6 GHz band wireless
communication links and dynamically selects the best link based on wireless
status. We evaluate network performance in-the-field experiments for V2R
communication. From the result, our design could perform several times of
higher efficiency for data delivering which is compared with the similar
approach based on 2.4 GHz WLAN.
1 Introduction
Intelligent Transport Systems (ITS) for road transport realizes safer driving which is
under consideration of road traffic or road environmental information. ITS allows to
perform road traffic surveillance with roadside unit (RSU) which has several types of
sensor such as temperature or CCTV. These collected information is realized for road
administrative operation and make road traffic safely providing information for drivers.
Another evolution of ITS can be identified for vehicle onboard sensor gets popu-
larly equipment. Some modern commercial available vehicles provide notification of
forward traffic signs for drivers, for example, when the vehicle is closing to stop sign at
the intersection, vehicle recognizes it using onboard camera and make notification for
driver. Onboard sensor allows to realize lane departure warning or collision avoidance
functions, this function is categorized into SAE’s 1 or 2 autonomy level [1]. In dis-
cussion for realization of higher level of autonomy vehicle, there is a considerable
challenge how the vehicle should collect forward road surface condition. Any auton-
omy vehicles should be considered of maneuvering which corresponds to road
https://doi.org/10.1007/978-3-030-15035-8_50
Network Performance Evaluation 525
condition in order to safely control even at snowy or icy road. Especially, in the cold
district, road condition often changes into short distance and in short time which is
caused by weather changing, thus it is required to collect actual road surface condition
in consecutive analyzing on sensing vehicle.
Due to above-mentioned methods which are based on onboard analyzing, the
required system functions must be supposed for system functions supposed for use of
wireless communication in effort to exchange with other vehicles. Floating car data
(FCD) is a one of practical V2X communication technology which depends on public
cellular network, it has characteristics for being blind zone to originate from out of
service area such rural or mountainous area, thus, this system should design inde-
pendent from public cellular network.
Vehicle-to-Everything (V2X) communication is a concept of Vehicle-to-Vehicle
(V2V) communications, Vehicle-to-Road (V2R), and other vehicular related commu-
nications together and makes data communication without cellular or other public
wireless access network possible between vehicle and others or RSUs. As a back-
ground, modern ITS uses dedicated short-range communication (DSRC) as known as
IEEE 802.11p standard for delivering data. However there is limitation of transmission
speed and communication range which are not enough when road condition
exchanging system increases amount of delivering data or V2V communication.
Another key technology to realize a connected car is low-power wide area (LPWA)
networking which provides long-range distance in several kilometers wireless com-
munication but provides quite low throughput of data delivering at a time.
This paper proposes a wireless communication system designed for V2X envi-
ronment in order to deliver condition of roads. In the principal design, the V2X
communication deals with associated road state such dried/wet/snowed/icy/whiteout
etc., and its precise location. The decision of road states is performed onboard node into
each vehicle by collecting and calculating of multiple sensor value, afterwards the
proposed system sends the result to roadside unit or other vehicle directly. In this paper,
we introduce three different wavelength wireless devices which contain of 2.4 GHz,
5.6 GHz and 920 MHz LPWA wireless network. In our design, 2.4/5.6 GHz wireless
LAN (WLAN) bands conduct data transmission mainly. Authentication and initial-
ization of connection on WLAN bands connection both will be sent over 920 MHz
band before vehicle is in WLAN range. The system design is designed for very high
speed and highly efficiency wireless network bandwidth in road condition data trans-
mission even vehicle is moving fast and road information becomes richer. Beam-
forming technology which is implemented 802.11ac standard allows us to realize to
realize more efficiency delivery of packets and to reach farther devices.
We have evaluated network performance to confirm our approach by field exper-
iment. The result indicates several times of higher efficiency for data delivering which
is compared with existing implementation based on 2.4 GHz WLAN standard.
526 A. Sakuraba et al.
2 Related Works
Road sensing technique processes in order to determine what road condition is cor-
responded to by analyzing of sensor output which is placed on moving vehicle. These
methods had been introduced in field of pavement engineering with non-contact
onboard sensor to collect pavement condition while probing vehicle is running. In this
method, dynamics sensors especially acceleration sensor is one of popularly sensor to
obtain road condition such pothole or bump of pavement. Chen et al. described their
road monitoring system to process sensor value for power spectral density in order to
estimate the International Roughness Index (IRI) using 3-axis acceleration sensor. As
an example for ideas based on using of non-dynamics sensor for road condition
sensing, Nakashima et al. discussed a method which is implemented with ultrasonic
sensor. Observation of quasi electrostatic field (QEF) between road surface and vehicle
body based road sensing method brought suggestion which was useful to determine it,
although the result showed that it required more accuracy method for determination of
road surface condition is equivalent to compacted snow and icy.
Almost practical implementations in order to provide floating car data function is
based on public wireless network such 3G/3.5G/3.9G/4G cellular network which is
available to deliver data in from several hundred kbps to dozen Mbps throughput.
Several automotive industries provide them in the market. For example, Honda
Internavi [2] allows precisely estimation of traffic jam which is jointly analyzed with
information from RSU and floating data from other vehicle. In their architecture, V2V
communication depended on public cellular network. CarTel [3] has V2R communi-
cation on private wireless network which is short-range vehicular adhoc network
(VANET) based on 802.11b WLAN, it was possible to deliver data which is capable
about 30 KiB/s end-to-end throughput from vehicle to internet.
We can cite a number of VANET proposal methods which is based on DSRC or
802.11p, and 802.11a/b/g/n for the popular wireless standard using for V2R or V2V
communication device. WAVE system [4] is a prototype of DSRC operated on
5.8 GHz band and designed to notify traffic signs for driver with V2R communication.
RSUs which equips the improved DSRC based wireless device are now in operation at
actual expressway of Japan. Bellavista et al. proposed a V2X communication protocol
for traffic light state exchanging method among other cars and RSU [5]. Their method
considered to operate in realistic scenario and congested urban area, it designed on
2.4 GHz 802.11b wireless with group-based communication protocol.
V2X applications especially require wireless communication short latency and less
packet loss ratio not only high efficiency data delivering, there is several implemen-
tations of cognitive radio for V2X communication which consists of multiple wireless
link and switches it situationally. Ito et al. developed 2-wavelength wireless commu-
nication consists of 920 MHz and 802.11b/g/n 2.4 GHz WLAN device to reduce
latency which is caused time of authentication and association which is performed
before delivering on WLAN band, they confirmed that the system allows to increase
amount of data delivering between vehicle and RSU [6].
3 Proposed System and Configuration
This section describes road condition sharing process of our proposal system which can
obtain sensor value and determine road condition on the vehicle, and then exchanging
with other vehicle and/or RSU and share among neighbor vehicles.
3.1 Overview
Figure 1 illustrates an overview of our system. Our proposed system can share with
highly real-time road states for drivers in manual driving.
Fig. 1. System overview of N-wavelength V2X wireless communication system
Onboard unit is loaded on each vehicle which consists of Sensor Server and SMB
(Smart Mobile Box). Sensor Server deals collecting of multiple onboard sensor output
firstly. SMB analyzes for decision road states of up to current travelling point which
describes as road state corresponding to dried/wet/snowy/icy/whiteout etc. with gran-
ularity. Next, SMB associates road state and their location. Then SMB delivers it with
raw sensor data to RSU called Smart Relay Shelter (SRS) on N-wavelength V2R
cognitive wireless network. On other side, roadside server provides road data and states
of wide area is collected by other vehicle via the same network. This cognitive network
also takes V2V communication directly with a vehicle which is heading to the opposite
direction from own vehicle. At the same time, SRS also perform to deliver regional
road states and acquires other area road surface states with cloudlet on cloud computing
resources, therefore our approach allows to provide service is public cellular network
independent.
Finally, the system provides visualization of road states based GIS (Geographic
Information System) for users of road such drivers or operation managers of public
transportation system such bus or taxi. For operation managers or driver who is not in
driving, our proposal system represents road states as generic GIS designed for using
on smart device and PC. Each vehicle equips road surface warning system in order to
represent and warn for driver who is in driving, if ahead road condition would get
worse which is described as icy, whiteout, etc. system will warn it in visually on
display device and playout of warning message by human voice.
Continually acquiring of road states information is required to provide precise
information, it could be efficient to load onboard unit on frequently running vehicles
along the same route such bus.
These functions allows acquiring, sharing, and representing of road condition, it
could realize safer road traffic even winter season.
3.2 Configuration of System

System configuration is shown in Fig. 2.
Fig. 2. System configuration
Sensor server monitors output value from sensor group which is composed of
various sensor of acceleration, gyro, geomagnetic, temperature, humidity, atmospheric
pressure, quasi electrostatic field (QEF), RGB image, and positioning of vehicle in
order to determine road surface state where vehicle is running. Sensor server is based
on edge computing architecture and processes acquired sensor values to identify road
state by machine learning algorithm. After road state decision has been completed, the
result of analyze is sent to SMB.
SMB has several wireless networks with different wavelength wireless network
device and decision maker which wireless device should be delivered to other SMB or
SRS.
At this stage we designed our proposal system has two of different wavelength
network device but extensibility is considered use of N-wavelength wireless link, such
IEEE 802.11ax or 802.11ay standard based one, into system design. Detail of using
wireless networks are described in Sect. 3.3.
3.3 Cognitive Wireless Network

Our proposed system has cognitive wireless network in order to perform data deliv-
ering with the best network throughput. The system consists of multiple network
devices which have different wavelength wireless links. In implementation at this time,
we install three different wireless standard based devices which is composed of
2.4/5.6 GHz band WLAN and 920 MHz band LPWA. We expect that system will be
able to provide better throughput for application by link selection which corresponds to
type of data/information.
2.4 GHz Band WLAN. 802.11b/g/n is the most popular wireless standard and used in
many field today, as well as VANET platform. 2.4 GHz band is a license-free band
both indoor and outdoor, and also allows to provide adhoc networking easily. How-
ever, as this wireless band would be encountered with terrible interference due to
carriers of other traffics, therefore 2.4 GHz band has concerned of system could not
expect much performance especially in urban area. In current implementation, we
designed the system to deliver road state information over 2.4 GHz WLAN which
could be composed of less amount of data.
5.6 GHz Band WLAN. We assumed that 5.6 GHz band is based on 802.11ac wave 2
standard, as primary data transmission link for exchange road surface states and raw
sensor data. 802.11ac wave 2 standard has capability for higher throughput compared
with existing 2.4 or 5 GHz WLAN standard, by multi-user multiple-input and multiple-
output (MU-MIMO) and channel bonding. MU-MIMO can deal multiple vehicle
communication at once without network performance degradation. Another charac-
teristics is beamforming technology, it can extend longer distance wireless commu-
nication with high throughput. Both characteristics are could be expected to provide
higher throughput even in longer range, it helps reliable and massive data transmission
while vehicle is moving faster. Our system uses 5.6 GHz band WLAN as the primary
link which delivers massive amount of data such raw sensor value or images of
onboard RGB sensor etc.
920 MHz Band LPWA. 920 MHz band wireless network, as known as LoRa or
LPWA, is one of license-free long distance radio communication technology. Network
performance of this type wireless technology is designed to perform several tens of kilo
bps throughput with several kilometers range distance, it is difficult to transmit large
amount of data as primary data transmission link but it is suitable to utilize it as
controlling message link. We designed our system which has common control channel
among SRSs and SMBs on LPWA long distance wireless network. This is a reason that
typical authentication of WLAN such Wi-Fi Protected Access II (WPA2) requires to
connect in several to several dozen seconds, vehicle easily moves to out of wireless
LAN coverage area while SMB is attempting to authenticate. This method could
increase amount of transmittable data than conventional single wavelength WLAN
configuration.
3.4 Data Transmission Flow in 2-Wavelength Network

Figure 3 describes data transmission while V2R and V2V communication using 2-
wavelength network devices from initiation to be connected to WLAN.
(1) Message & data flow in V2R (2) Message & data flow in V2V
Fig. 3. Transmission of data and message flow. (1) describes V2R transmission in the event of
destination vehicle requests information to SRS. (2) describes destination vehicle requests it to
source vehicle in adhoc wireless communication
Messaging flow of V2R scenario is illustrated as Fig. 3-(1). Initially, SRS broad-
casts own existence information via 920 MHz band, which corresponds to beacon
frame of access point in 802.11 based WLAN. This information includes traffic
direction which SRS collects and provides road states. When running vehicle receives
broadcast message from SRS which is installed in the same traffic direction as the
vehicle is moving, SMB responses to SRS with own node ID. Responding to that, SRS
assigns an IP address for WLAN on SMB and then deliver WLAN network infor-
mation for SMB. These information is composed of assigned IP address, netmask,
default gateway address, ESSID of access point, WPA passphrase etc. due to SRS
requests SMB to connect Infrastructure Mode in V2R communication. SMB activates
WLAN device after connection information has been received and waits for vehicle is
in range of WLAN of SRS. When the vehicle entered into WLAN coverage area, SMB
attempts to take 4-way handshake for WLAN connection. If WPA authentication has
been succeeded, it also checks for continuity by ICMP message exchanging. At this
point WLAN connection has been established completely, thus both vehicle and SRS
are able to exchange road state information over TCP each other.
Figure 3-(2) denotes message flow which describes V2V communication from
initiation to be established. Message and data flow is very similar procedures, firstly
destination vehicle which requires road state information from other vehicle, waits
broadcasts own existence information from source vehicle which has already road state
information, via 920 MHz band. If SMB on destination vehicle received broadcast
message, sends node ID in response. In V2V communication, source vehicle answers

IP address related information, ESSID, and wireless-key to destination vehicle. After
destination vehicle received these information, it activates WLAN adapter and attempts
to connect in adhoc mode on WLAN until both vehicles in range of it. After WLAN
connection has been established, destination vehicle measures for continuity, finally
road condition exchange will be taken over TCP in adhoc mode between both vehicles
if continuity test has been passed.
4 Prototype Environment
We are now implementing three-wavelength onboard/roadside communication device

as the prototype system.
We built three-wavelength wireless communication device which uses Buffalo UI-
U2-300D for 802.11b/g/n WLAN USB device, Planex communication GW-900D
USB WLAN adapter for 802.11ac WLAN, and Oi Electric OiNET-923 for 920 MHz
LWPA wireless network, these devices are connected to SMB.
SRS also has the similar configuration but Ruckus Wireless Zoneflex T300 access
point for 802.11ac WLAN capable device instead of GW-900D.
Controlling of SMB and SRS is based on AMD64 architecture desktop system, we
implemented them on Intel NUC Kit NUC5i5RYH barebone PC kit with Ubuntu 16.04
Linux system. Application on SRS and SMB is written by C, Ruby, and Bash Script.
5 Evaluation
We had an experiment of the V2R network based performance on our N-wavelength

wireless system to evaluate availability in actual roadside environment. In this stage,
system has two-wavelength wireless system which contains 920 MHz/5.6 GHz band
wireless links. In this section, we describe this performance evaluation and consider
future challenges.
5.1 Configuration and Network Setup

Figure 4 illustrates configuration and network setup for performance evaluation. We
placed a pair of cognitive wireless unit which consists of 920 MHz LPWA and
802.11ac access point (AP) for SRS and 802.11ac USB wireless device on SMB.
802.11ac link is capable for IP networking, but LPWA is not based on it.
We set scenario which is denoted as when SMB uploads accumulated road
information and sensor data to SRS. In this scenario, the SMB and SRS exchanges
network settings information on 920 MHz link, and then SMB sends a file which is
sized 400 MiB over FTP on 2.4/5.6 GHz link.
We compare our proposed method, generic 802.11ac network with DHCP and
WPA2, and Ito et al.’s work [6] which uses 802.11b/g/n wireless standard. As shown in
Fig. 4, we added 2.4 GHz wireless device to both units in experimental system in order
to operate 2-wavelength cognitive wireless network which is implemented in previous
work.
Fig. 4. Device connection and configuration of experimental setting on SRS and SMB
5.2 Environment for In-the-Field Experiment

We have experimented to evaluate our method on actual vehicle. We set two scenarios
to measure performance difference which is caused moving direction of vehicle. The
scenario requires vehicle to move to eastbound or westbound direction as be described
in Fig. 5. Vehicle moves at 20 km/h constant speed on either direction of the road. We
installed both 2.4 GHz and 5.6 GHz band WLAN unit on the top of left side view
mirror on the vehicle.
Fig. 5. Overview of both vehicle direction trials and installation of SMB wireless units
We had planned this evaluation in actual public two-lane road which is illustrated
in Fig. 6. The length of evaluation section on road was 1.1 km and road alignment was
almost straight and almost smooth terrain but there was several grade section which
includes 3.3% grade in 90 m length. We placed SRS on sidewalk where is located next
to the westbound lane. The radio antenna was fixed on an antenna pole and their height
Fig. 6. Road environment and SRS placement point in the field
from the ground was 2.2 m for LPWA, 1.7 m for 802.11ac, and 1.4 m for 802.11b/g/n.
802.11ac WLAN unit on SRS faced to direction across of road in 90°.
The condition of road traffic density was medium and weather was snowfall.
5.3 Measuring Target

The system recorded several types of network states which includes received signal
strength on each wireless links, the necessary time to connect WLAN from system
initiation, and delivered data amount. In this paper, we discuss them partially, espe-
cially delivered data amount, in order to evaluate our current work.
5.4 Result
We have ran 5 trials each three different types configuration and both vehicle’s
directions. Figure 7 plots delivered data amount which is succeeded to send from SMB.
Generically, 802.11ac link can deliver more efficiently in every trials. The previous
method which uses 2.4 GHz link was confined to deliver averagely sent 53.1 MiB
(s.d. = 8.5) in eastbound and 52.8 MiB (s.d. = 7.1) in westbound trials, nevertheless
our system averagely delivered 218 MiB (s.d. = 109.7) from moving vehicle to SRS in
eastbound and 301 MiB (s.d. = 65.8) in westbound. The result indicates that our
method was 4.1 to 5.7 times faster than 802.11b/g/n based system.
We also measured delivered packet in environment which based on generic
802.11ac, the averaged delivered data was 125 MiB (s.d. = 175.2) in eastbound trials
and 259 MiB (s.d. = 160.0) westbound one. The result recorded three times of data
delivering failure in generic 802.11ac setup and it contains two times timeout while
FTP session establishing, and once authentication failure on WPA2 before delivering
IP packets.
5.5 Discussion
The result proved that high-speed wireless link even short distance coverage, such
802.11ac which is a sort of license-free bands, could be allowed to deliver a large
(1) Sent data amount to SRS during 5 trials when (2) Sent data amount to SRS during 5 trials when
vehicle moves to west vehicle moves to east
Fig. 7. Plotted delivered data amount while V2R data exchanging over each wireless system
amount data from moving vehicle to SRS. For example, it is capable to exchange long
time RGB sensor output which corresponds to in 10 min recording if averaged size of
compressed road surface image equals to 500 KiB and it is recorded in a frame per
second. We suggest reason of availability for large amount data delivering was caused
of beamforming technology on 802.11ac standard has an effect on high throughput
which was observed in our field experiment.
The effect for packet delivering performance improvement on 5.6 GHz band with
920 MHz LPWA is not able to be concluded in this time, due to unstable performance.
Thereby we are required to repeatedly examine both network performance in the field.
Another topic for discussion is performance difference which is depended on
relationship between SRS placed location and traveling lane of the vehicle was contact
with SRS or not. In our result, westbound trials demonstrated 38% better performance
which compared with eastbound trials for delivering packet. This result could be
attributed to multiple factors such as oncoming vehicle, objects around circumference
of road such building, trees, etc. Especially, we surmised the distance among onboard
WLAN units and SRS in westbound trials which was clearly shorter than eastbound.
This result suggests to install multiple WLAN units on vehicles such right, left, and rear
of the body of car, in effort to obtain stronger signal between SRS and SMB, therefore
it realize better network performance.
6 Conclusion
This work proposed a design for N-wavelength wireless communication system for
V2X environment to exchange road surface conditions. This system has three of
fundamental functions; observing of physically state of road surface by multiple sen-
sors on the vehicle, analyze these sensor output for decision of road surface conditions,
and transmit calculated road condition to roadside device or other vehicle to share road
condition which the vehicle travelled until now. We designed that this road surface
state provides to warn drivers both own and others if ahead road surface condition will
be getting worse. Road condition is also provided in road service GIS for person who is
not a driver.
Our system uses various different wavelength wireless networks among vehicles
and RSUs which is not depended on public wireless WAN, in order to be available the
functions even mountainous district. Our current work is implementing with two-
wavelength wireless network consists of 920 MHz and 2.4/5.6 GHz license-free bands.
The system delivers road state information over 2.4 GHz, send raw sensor values over
5.6 GHz, and exchanges network or authentication information for WLAN over
920 MHz LPWA band, this design allows onboard cognitive wireless communication
unit to immediately send packet from WLAN devices when vehicle will be just
entering into WLAN coverage of SRS. This design is considered to improve packet
delivery efficiency for communication among moving vehicles.
We have measured V2R communication performance in actual road environment.
The result of evaluation proved that our proposed approach is 4.1 to 5.7 times faster
than previous approach which is based on 2.4 GHz WLAN. However effect on network
performance improvement using 920 MHz LPWA was not clear. The result also
suggested to install multiple WLAN unit on the each sides of vehicle.
In our future work, we are considering a new networking protocol to realize the
most efficiency on WLAN band which is able to packet delivering less than a second
when vehicle moved into range of WLAN. We are now investigating of combination of
LPWA and FILS (fast initial link setup) which is standardized as IEEE 802.11ai [7].
Acknowledgments. This research was supported by Strategic Information and Communications

R&D Promotion Program (SCOPE) No. 181502003, Ministry of Internal Affairs and Commu-
nications, Japan.
References
1. SAE International: Taxonomy and Definitions for Terms Related to Driving Automation
Systems for On-Road Motor Vehicles J3016_201806 (2018)
2. Honda Global, October 17, 2003 Honda Introduces World’s First Congestion Prediction
Function for it’s InterNavi Premium Club Car Navigation Service. https://global.honda/
newsroom/news/2003/4031017-navi-eng.html. Accessed 20 Jan 2019
3. Hull, B., et al.: CarTel: a distributed mobile sensor computing system. In: Proceedings of the
4th Conference on Embedded Networked Sensor Systems (2006)
4. Tsuboi, T., Yamada, J., Yamauchi, N., Hayashi, M.: Dual receiver communication system for
DSRC. In: Proceeding of 2008 2nd International Conference on Future Generation
Communication and Networking, pp. 459–464 (2008)
5. Bellavista, P., et al.: V2X protocols for low-penetration-rate and cooperative traffic
estimations. In: Proceedings of 2014 IEEE 80th Vehicular Technology Conference,
pp. 1–6 (2014)
6. Ito, K., Hashimoto, K., Shibata, Y.: V2X communication system for sharing road alert
information using cognitive network. In: Proceeding of 8th International Conference on
Awareness Science and Technology, pp. 533–538 (2017)
7. IEEE 802.11ai-2016 - IEEE Standard for Information technology - Telecommunications and
information exchange between systems - Local and metropolitan area networks - Specific
requirements Part 11: Wireless LAN Medium Access Control (MAC) and Physical Layer
(PHY) Specifications Amendment 1: Fast Initial Link Setup (2016)
A IoT Based Disaster Information Platform
for Challenged Network Environment
in Snow Countries
Yoshitaka Shibata(&), Yoshikazu Arai, Yoshia Saito, and Jun Hakura
Iwate Prefectural University, Takizawa, Japan

{shibata,arai,y-saito,hakura}@iwate-pu.ac.jp
Abstract. In many snow countries, since the traffic information and commu-
nication networks infrastructures are not well developed due to their inherent
weather and geological conditions and rapid progress of aging and depopulation,
the resident’s safe livings are always suffered from those challenged environ-
ments, such as disasters, heavy snow, iced road conditions. In order to release
from those serious problems, a new IoT based information platform is intro-
duced in this paper. On the proposed platform, mobility such as cars and drones
can performs information and communication means to organize an information
infrastructure in challenged network environment where both wired and wireless
network are not well installed such as mountain areas and coast areas in snow
countries. In the disaster cases as emergency situations, drones are launched
from cars and can organize temporal wireless mesh networks in the sky and can
cover the wide communication areas. Thus by combing both the cars and
drones, very important information related resident’s life and property can
quickly and flexibly exchanged using this platform. In this paper, the general
system configuration and architecture and its expected application for disasters
are discussed.
1 Introduction
Many areas in Japan is occupied by mountains and forests. However, in those areas,
aging population, depopulation and declining economic activity are rapidly advancing
[1, 2]. Particularly in the cold areas, the residents are forced to drive in the chronic
serious bad road conditions, such as snow storm, frozen road, snow slide, whiteout. For
this reason, the number of traffic accidents are increasing as the aging and abnormal
weather conditions are advancing. The information and communication environment
along the roads in local areas are not well developed compared with urban areas. Once
a traffic accidents or disaster occurred, information collection, transmission and sharing
are delayed or even cannot be made. Eventually the resident’s lives and reliabilities
cannot be maintained. More proper and quick information service system with disaster
and road environmental states are indispensable.
In order to resolve those problems, we propose IoT based disaster information
platform. In our system, road side wireless nodes and mobile smart nodes with various
sensors and different wireless communication devices organize a largescale information

https://doi.org/10.1007/978-3-030-15035-8_51
538 Y. Shibata et al.
infrastructure without conventional wired network. The mobile nodes such as cars and
drones collect various sensor data and carry those data as a message ferry while moving
one end to another along the roads.
On other hand, the road side wireless node not only collects and stored sensor data
from its own sensors in its database server but exchanges the sensor data from the
mobile nodes when it passes through the road side wireless node by V2X protocol [3].
Therefore, both sensor data at the road side wireless node and the mobile nodes are
periodically uploaded to cloud system and synchronized. Thus, mobile node perform as
mobile communication means even through the communication infrastructure is
challenged environment or not well developed.
This disaster information platform not only performs various road sensor data
collection and transmission functions, but also performs Internet access network
functions to transmit the various data such as disaster prevention information on
emergency case, road state information, sightseeing information and shopping infor-
mation on ordinal case. Therefore, many applications and services can be realized.
2 Disaster Information Platform
Since the population density in the mountain or rural area is quite low, the commu-
nication infrastructure is not well developed although many residents in those areas
spend daily life. Therefore, it is desired to realize information infrastructure to take
advantage of automobile. Automobile can always move anywhere along roads and
encounter the other automobile. Also by installing the base stations along roads, the
automobile can communicate and exchange to those base stations and eventually
communicate to Internet and cloud computing. Furthermore, even though the road
network infrastructure is well developed, drone can fully or partially perform the
function of automobile. In the future, since the drone improves its flying performance
and expected to launch up from and arrive directly to automobile, more flexible and on
demand information exchange functions can realized.
The Fig. 1 shows a system configuration of our proposed IoT based mobility
information infrastructure which consists of multiple wireless networks and nodes
including fixed road side base station, balloon, mobile vehicle, drone in addition to
satellite and various offices and shelters. Those nodes are mutually connected with
various wireless communication networks including LPWA, 3G, LTE, Wi-Fi, Wi-Max,
satellite in addition to wired networks. In the future, 5G is also connected as wireless
network. The mobile wireless nodes are organized by different typed wireless network
devices and various sensor devices such as camera, semi-electrostatic field sensor,
acceleration sensor, gyro sensor, temperature sensor, humidity sensor, infrared sensor
and sensor server.
A IoT Based Disaster Information Platform 539
Fig. 1. IoT based disaster information platform
3 Basic Functions of Platform
Figure 2 shows a fixed wireless node including fixed road side base station, various
offices and shelters which performs sensing environmental conditions such as road
condition, mountain condition and disaster area state using camera, temperature,
humid, rain, wind power, etc. Those fixed nodes also communication function, storage
function of the sensing data and exchange function of the sensor data from the mobile
node and autonomous power supply system which is consisted of window turbine,
solar panel, battery and controller [10, 11, 13].
Fig. 2. Fixed wireless node
As wireless network communication means, multiple wireless network devices with

different frequency band, such as IEEE 802.11g (2.4 GHz), IEEE 802.11ac (5.6 GHz),
LPWA or IEEE 802.11p (920 MHz), IP satellite IP network in addition to 3G, LTE,
Wi-MAX are used and organizes a cognitive wireless node. The network node selects
the best link of cognitive wireless network depending on the observed network quality
such as RSSI, delay, packet loss rate using Software Defined Network (SDN) function
[12]. If none of link connection is existed, those transmitted sensor data and disaster
information are locally and temporally stored in the storage until another mobile node
approaches, and starts to transmit data by Delay Tolerant Network (DTN) function [4–
6]. Thus, data communication can be attained even though the network infrastructure is
destroyed or not existed in challenged network environment such as mountain areas or
just after large scale disaster areas.
Figure 3 shows mobile wireless node including automobile, drone and balloon
nodes and consists of various environmental sensors, cognitive wireless network with
different frequency bands as the fixed wireless nodes, Software Defined Network
(SDN) function, and Delay Tolerant Network (DTN) function and Storage function.
Fig. 3. Mobile wireless node
Various mobile sensors including accelerator, gyro sensor, infrared temperature

sensor, humidity sensor, quasi electrical static sensor, camera and GPS are integrated to
precisely and quantitatively detect the various road surface and surround states and
determine the dangerous locations on GIS in sensor server. Those data and information
are mutually exchanged to the other automobile and the road side base station.
4 Drone Based Communication

4.1 Drone to Vehicle/Base Station Communication
In order to monitor the wide state in mountain area in a short time than only automobile
communication, multiple drones are used. This is very effective to know the state of
fire, avalanche, landslide, eruption as well as to find the victims, wild animals. In
addition to those monitoring functions, data communication and exchanging can per-
formed when vehicles cannot pass the roads due to disaster. By launching the drones
from automobile and monitoring the objective areas by camera and transmitting those
data to the base stations and or automobiles on the ground through cognitive wireless
network, the state of the mountain area, roads and rivers can be widely monitored. In
the case of disconnecting between automobile and base station, the drones temporally
accumulate those data to the storage using DTN protocol. When the possible
communication began to connect to neighbor automobile and base station, the best
communication link is selected using SDN function and eventually those accumulated
data are downloaded to the communication node on the ground.
4.2 Drove to Drone Communication

In order to realize the case of the communication in the mountain area or on the
seashore where automobile cannot run nor communicate each other, multiple drones
can organize a wide are mesh network in the sky and communicate and exchange
information each other or the vehicles and base stations. In order to attain more
effective communication and exchanging the information, the previous flight pass
information are summarized to determine the best location that drones can encounter
with the highest probability and exchange information each other in the shortest time
(Fig. 4).
Fig. 4. Drone to drone/vehicle communication
5 Data Migration and Coordination
All of the collected data from the sensors are linked to the equivalent the geological tab
such as latitude, longitude, height and time. However, those data are not temporally and
spatially homogeneous because the sensor locations are irregular position nor the
sampling period and duration are not well synchronized. In another words, those data
have both geologically sparse and temporally time series characteristics. Therefore, in
order to process those sensor data to make a decision for the state of environmental
state. In our research, more active measuring method is introduced in which those
nonhomogeneous data can be compensated to the homogeneous data by temporal and
spatial interpolation and extrapolation function and integration and distribution func-
tion by optimizing the sensing granularity in both spatial and time domain.
6 Disaster Application
The Figs. 5 and 6 show a Tourist and Disaster Information using our proposed system
for normal case and emergent case [7–10]. In normal situation, many Japanese tourists
and foreign tourists from different countries visit to the Points of Interest (POI) such as
the historical places, national parks and mountains and spas where network information
infrastructures are usually not well prepared in snow country in Japan. Even worse, it is
difficult for foreigners to understand the information of Points POI written in Japanese.
However, by introducing the proposed mobility information infrastructure, and using
smart phones and tablet terminals, the tourists can ask and speak by their own voices
and obtain information POI by their own languages, eventually being navigated to the
objective places.
Fig. 5. Tourist information delivery on normal case
On the other hand, when disaster occurred, the push typed disaster information,
evaluation information are automatically delivered to those tourists by their languages.
Eventually, the tourist can safely evacuate to the proper shelter from the current
location. Through the mobility information infrastructure, the disaster state informa-
tion, resident safety information, required medicine, feeds and materials are also col-
lected and transmitted by mobile nodes between the counter disaster headquarter and
evacuation shelters as shown in Fig. 6.
Fig. 6. Disaster information delivery on emergency case
7 Conclusions
In this paper, a new IoT based disaster information platform is introduced. On the
proposed platform, mobility such as cars and drones can performs information and
communication means to organize an information infrastructure in challenged network
environment where both wired and wireless network are not well installed such as
mountain areas and coast areas in snow countries. In the disaster cases as emergency
situations, drones are launched from cars and can organize temporal wireless mesh
networks in the sky and can cover the wide communication areas. Thus by combing
both the cars and drones, very important information related to resident’s life and
property can quickly and flexibly exchanged using this platform. Currently we are
implementing a prototype of the proposed the disaster information platform by
extending previously developed Never Die Network (NDN) to evaluate the effects and
usefulness of the proposed platform through annual disaster evaluation drill in a snow
country.
Acknowledgement. The research was supported by Strategic Information and Communications

R&D Promotion Program Grant Number 181502003 by Ministry of Affairs and Communication.
References
1. A 2016 Aging Society White Paper. http://www8.cao.go.jp/kourei/whitepaper/w-014/
zenbun/s1_1_1.html
2. A 2016 Declining Birth rate White Paper. http://www8.cao.go.jp/shoushi/shoushika/
whitepaper/measures/english/w-2016/index.html
3. Ito, K., Hirakawa, G., Shibata, Y.: Experimentation of V2X communication in real
environment for road alert information sharing system. In: IEEE AINA 2015, pp. 711–716
(2015)
4. Otomo, M., Sato, G., Shibata, Y.: In-vehicle cloudlet computing based delay tolerant
network protocol for disaster information system. In: Advances on Broad-Band Wireless
Computing, Communication and Application Applications. Lecture Notes on Data
Engineering and Communications Technologies, vol. 2, pp. 255–266 (2016)
5. Hirakawa, G., Uchida, N., Arai, Y., Shibata, Y.: Application of DTN to the vehicle sensor
platform CoMoSe. In: WAINA 2015, pp. 24–27 (2015)
6. Kitada, S., Sato, G., Shibata, Y.: A DTN based multi-hop network for disaster information
transmission by smart devices. In: Advances on Broad-Band Wireless Computing,
Communication and Application Applications. Lecture Notes on Data Engineering and
Communications Technologies, vol. 2, pp. 601–611 (2016)
7. Shibata, Y., Goto, T., Sato, G., Hashimoto, K.: Disaster information sharing system
considering communication status and elapsed time. In: IWDENS (2017)
8. Shibata, Y., Uchida, N., Shiratori, N.: Analysis and proposal of disaster information network
from experience of the great East Japan earthquake. IEEE Commun. Mag. 52, 44–48 (2014)
9. Japan Police Department: The Great East Japan Disaster. http://www.npa.go.jp/archive/
keibi/biki/index.htm
10. Uchida, N.: Min-max based AHP method for route selection in cognitive wireless network.
In: Proceedings of the 2010 13th International Conference on Network-Based Information
Systems, NBIS 2010, pp. 22–27 (2010)
11. Uchida, N., Takahata, K., Shibata, Y.: Disaster information system from communication
traffic analysis and connectivity (quick report from Japan Earthquake and Tsunami on March
11th, 2011). In: The 14th International Conference on Network-Based Information Systems
(2011)
12. Open Networking Foundation: Intro to OpenFlow. https://www.opennetworking.org/
standards/intro-to-openflow
13. Uchida, N., Sato, G., Shibata, Y., et al.: Selective routing protocol for cognitive wireless
networks based on user’s policy. In: The 12th International Workshop on Multimedia
Network Systems and Applications (MNSA 2010), pp. 112–117 (2010)
Study on Balloon Network Using LoRa Mesh
Communication System
Goshi Sato1(&), Yoshitaka Shibata2, and Noriki Uchida3

1
National Institute of Information and Communications Technology,
4-2-1 Norikibe-cho, Koganei, Tokyo 184-8795, Japan
sato_g@nict.go.jp
2
Faculty of Software and Information Science, Iwate Prefectural University,
152-52 Sugo, Takizawa, Iwate 020-0693, Japan
3
Higashi-Ku, Fukuoka 811–0295, Japan
n-uchida@fit.ac.jp
Abstract. LoRa wireless communication needs to be installed as high as

possible in order to fully demonstrate its communication distance. However, in
times of disasters, there are not necessarily high structures that can be conve-
niently installed. Therefore, in this research, by attaching the LoRa mesh net-
work system developed by NICT to helium balloon and drone, it will make
network emergency restoration more realistic in the event of a disaster. We aim
to operate 72 h after the occurrence of a disaster, and promote power saving and
weight saving.
1 Introduction and Purpose
LoRa wireless communication is being used in recent years as a wireless communi-

cation standard for IoT capable of long-distance communication with low power
consumption. We have confirmed through experiments that the communication dis-
tance of the LoRa radio and the installation location are very closely related.
In the experiment carried out in Kamaishi city, Iwate prefecture, we attempted
communication by placing the LoRa radio at about 2 m above the ground as shown in
Fig. 1. As a result, it was confirmed that packet loss occurs even at a point of 1 km or
3.5 km depending on parameters to be set.
Naturally, it is important how high you can set up the radio. However, the LoRa
mesh network system dealt with in this research is easy to install, and it is superior to
automatically form a network by simply turning on the power, but in this case
installation location becomes a problem. Also, in the event of a disaster, there is not
always a high place that can be conveniently set up at the base, so in order to maximize
the performance of the LoRa mesh network system, it is necessary to build a sky
communication network with balloons, drone, etc.

https://doi.org/10.1007/978-3-030-15035-8_52
546 G. Sato et al.
Fig. 1. Ground-to-ground LoRa communication experiments carried out across the bay
2 Methods
Communication by LoRa is characterized by being able to receive signals even with

received signal strengths below the noise level due to spreading gain in the modulation
scheme. Therefore, communication distance can be communicated with a communi-
cation partner more than tens of kilometers away if the antenna height is high and the
prospect can be obtained. There is a method of using a helium gas filled type balloon
when securing over-the-air communication by means not requiring a license or the like.
In this research, we develop a lightweight and power-saving balloon network system
based on LoRa mesh network technology as shown in Fig. 2 and implement appli-
cations to share the minimum necessary safety information among bases.
Study on Balloon Network 547
Fig. 2. Outline of construction of LoRa communication network by balloon
For situations like this, where we assume a network connecting centers that are
some distance away from each other, such as office space and evacuation centers and
exchange relatively lightweight data such as safety information, we set the spreading
factor and used channel bandwidth setting to slow But adjust to parameters that allow
longer distance communication.
Evaluate the procedures and construction time of the over-the-air network, human
resource, continuous operation time, and clarify the difference of features of ground
communication and over-the-air communication, advantages and disadvantages.
Measure the effective throughput of the constructed network and the time taken for data
synchronization and evaluate whether it meets the requirements of the application to be
implemented. Since LoRa wireless communication uses broadcast communication, if
collision of wireless signals occurs due to simultaneous transmission, retransmission is
not performed. Also, in consideration of various communication environments, nodes
that perform communication have hidden terminal problems because radio wave reach
ranges are extremely different.
In order to avoid simultaneous transmission, LoRa mesh network technology
implements time-division access control method after time synchronization by GPS. It
is assumed that all the radio stations operated in the system are numbered by ID and the
total number is also grasped. In this premise, based on the ID dynamically allocate time
slots that it can transmit. For example, assuming that the transmission time per wireless
device is 10 s (considering the time required for backoff to avoid collisions with other
systems), when using five wireless stations, these five devices do not have any prob-
lems A time longer than 50 s that can be transmitted is assigned as a transmission
cycle. In this transmission cycle, as shown in Fig. 3, transmission time slots are
assigned based on respective IDs. As a result, occurrence of collision in communication
in the own system is avoided. The amount of data that can be transmitted per unit time
548 G. Sato et al.
of the wireless device and the communicable distance vary depending on factors such
as the spreading factor of LoRa, the used channel bandwidth, the surrounding com-
munication environment, and the like. Also, since the transmission cycle varies
according to the number of radio devices used in the system, slot design of the entire
system is required in advance according to the requirements.
Fig. 3. Relationship between transmission cycle time and transmission time per wireless device
Therefore, in this research, by appropriately designing the spreading factor of the

LoRa, the used channel bandwidth, and the transmission cycle according to the fol-
lowing three kinds of application systems on the basis of the LoRa mesh network
technology, Verify feasibility. By the way, data relaying is done by using flooding.
However, in addition to the mechanism to avoid transmitting the same information
multiple times, in order to reduce the number of flooding transmission, the following
implementation is carried out. First, by sending information listing the sender IDs of
the signals received by each node in the past as control signals at regular intervals, it is
possible to grasp the range within which their neighboring nodes can communicate.
When a newly received information is received, if there is a node that advertised the
information earlier before its own transmission timeslot comes around, if the neigh-
boring node that it covers is adjacent to If it is all covered, it will not retransmit. As a
result, retransmission of unnecessary information is suppressed.
3 Experimental Environment Preparation
We are prototyping a prototype system (Figs. 4 and 5) that currently implements this
protocol based on NerveNet [1] developed by NICT, and we are carrying out a lab-
oratory test. Basic protocol operation can be confirmed by the expected operation, basic
data collection, system design, and implementation for each module are expected to be
completed in the current fiscal year. We will conduct information transfer experiments
using various data in future and demonstrate communication outdoors.
Study on Balloon Network 549
Fig. 4. Prototype system appearance Fig. 5. Prototype system internal image
4 Summary
In this paper, We proposed combining LoRa mesh network system developed by NICT
with overhead communication utilizing balloon and drone. I explained the basic
architecture of the LoRa mesh network system and outlined the balloon network for
applying it. We are also developing a power saving and lightweight wireless module
that can be attached to a balloon as a prototype system. In the future, we will form a
network between wireless modules actually mounted on the balloon and verify how
useful it is compared to ground communication.
Reference
1. Inoue, M., Ohnishi, M., Peng, C., Li, C., Owada, Y.: NerveNet: a regional platform network
for context-aware services with sensors and actuators. IEICE Trans. Commun. E94-B(3),
618–629 (2011)
Performance Evaluations on Adaptive
Array Antenna of Vehicular Delay
Tolerant Networking for Winter Road
Surveillance Systems
Noriki Uchida1(&), Goshi Sato2, and Yoshitaka Shibata3

1
Fukuoka Institute of Technology, 3-30-1 Wajirohigashi,
Higashi-ku, Fukuoka 811-0214, Japan
n-uchida@fit.ac.jp
2
Resilient ICT Research Center, National Institute of Information
and Communications Technology, 2-1-3 Katahira, Aoba-ku,
Sendai 980-0812, Japan
Sato-g@nict.go.jp
3
Iwate Prefectural University, 152-52 Sugo, Takizawa, Iwate 020-0693, Japan
Shibata@iwate-pu.ac.jp
Abstract. With the developments of ITS technologies such as V2V networks,

it is expected to realize the new kinds of applications such as the road
surveillance system during the winter. However, there are some limitations such
as the transmission ranges or the radio noises in the use of the higher frequency.
Therefore, this paper proposes the optimal AAA control methods with the
camera image recognition for the DTN method in the road surveillance system,
and it is especially discussed about the performance improvements of the remote
voltage controls by the asynchronous VISA commands in order to realize the
quick adjustments of the beam-forming for the target vehicles. Then, the
experiments by the prototype systems are presented in this paper, and the results
show the effectiveness of the proposed AAA controls.
1 Introduction
With the developments of ITS technologies such as V2V (Vehicle-to-Vehicle) net-

works, it is expected to realize the new kinds of applications such as the road
surveillance system during the winter. However, there are some limitations in the use of
the higher frequency for the V2V networks such as 5.9 GHz in the IEEE802.11p
standard. One of the limitations is the transmission range between the vehicles.
According to the specification of IEEE802.11p [1], the output power is 10dBm, and so
it is expected that the transmission range will be about a couple of hundred meters
between the vehicles. Therefore, it is necessary to consider the function of the direc-
tional antenna in order to make the wider transmission range. Besides, it is also nec-
essary to consider the quick movements of the vehicles and the radio noises from the
obstacles for the V2V communication, and so the adjustment of the radio directional
controls could be significant for the use in the field circumstances.

https://doi.org/10.1007/978-3-030-15035-8_53
Performance Evaluations on Adaptive Array Antenna 551
Thus, this paper proposes the optimal AAA (Adaptive Array Antenna) control
methods with the camera image recognition for the DTN (Delay Tolerant Networks)
method in the road surveillance system in the winter [2]. Also, it is especially discussed
about the performance evaluations of the remote voltage control methods for the
antenna elements in order to realize the quick adjustments of the beam-forming for the
target vehicles in this paper. The proposed system consists of the camera recognition
for the target automobiles, the beam-forming antenna that is calculated by the Kalman
Filter Algorithm [3, 4] in order to reduce the radio noises, and the voltage control
device for the antenna elements [5]. Then, this paper newly reports the improved
methods and the experimental results of the rapid voltage control methods for the
Adaptive Array Antenna (AAA) with the asynchronous communication by using the
VISA (Virtual Instrument Software Architecture) standard, and the future research
subjects are discussed.
In the followings, the proposed methods of the AAA with the image recognition for
the winter road surveillance system are introduced, and the prototype system in the
previous study is explained in Sect. 2. Moreover, the proposed rapid voltage controls with
the asynchronous communication by using the VISA are discussed in Sect. 3. Finally, the
experiments for the evaluation of the proposed AAA control methods are discussed in
Sect. 4, and the conclusions and the future researches are discussed in Sect. 5.
2 Proposed Systems
Figure 1 shows the proposed road surveillance system in the winter [2]. In the system,
each vehicle equips the wireless networks such as IEEE802.11p [6], and the camera
system to recognize the target vehicles, and the AAA system to adjust the beam-
forming radio in order to establish the longer transmission of the target vehicles. The
V2V networks are based on the DTN method since the proposed system is assumed to
use in the mountain area where LTE or 3G networks are not provided, and so the
vehicles are possible to receive the road information from the target vehicles through
the V2V communication. It is also assumed that the road information is acquired from
the sensors from each vehicle’s wheel or on each vehicle.
Fig. 1. The proposed systems for the road surveillance system in the winter.
However, it is necessary to consider the obstacles between the vehicles such as

trees or houses in the assumed assumptions. Here, if there is a tree between the node A
and the node B in Fig. 1, it is hard to present the stable connections in the system.
Therefore, the proposed methods include the predictional adjustments of the target
angle is confirmed by the Kalman Filter algorithm [6, 7] as shown in Fig. 2. Figure 2
shows the purposes of the proposed connectivity in comparison with the non-
directional antenna and the proposed AAA system. The proposed system realizes the
stable connections without the effects of the obstacles with the Kalman Filter Algorithm
and with more extended wireless connections with the beam-forming of the proposed
AAA system.
Fig. 2. The target connectivity of the DTN based V2V networks in the paper.
Fig. 3. The prototype system for the winter road surveillance system.
Also, since the rapid controls of the beam-forming, the proposed system introduced
the image recognition for the target vehicles. The previous papers [7, 8] introduced the
AAA controls by the MMSE (Minimum Mean Square Error) or the LMS (Least Mean
Square), but there are the problems of the convergence the complexity for the
calculation. Therefore, the AAA controls with the image recognition with the Machine
Learning algorithm is introduced in this paper, and the Kalman Filter algorithm is
introduced the reduction of the radio noises from the rapid movements of the auto-
mobiles on the roads.
Besides, the DTN routing [9, 10] is introduced in the proposed network commu-
nications. The DTN is the stored-and-carried typed protocols for the robust network
conditions such as the interplanetary communications originally, and it is necessary to
consider the mountain areas in the proposed systems. The previous study [11] indicated
that the prototype system of the proposed is shown in Fig. 3. The system mainly
consists of three parts such as the camera system (Logicool Web camera c270) for the
recognition of the target vehicle, the PC (Macbook 12inch) for the calculation of the
distance and the angle for the target vehicle and for the predictions by the Kalman
Filter, and the AAA system that is controlled by the voltage machine (PMX18-5A by
KIKUSUI Corporation). However, there were some problems with the performance of
the directional controls of the AAA system, and this paper will discuss the enhanced
voltage control methods by asynchronous communication.
3 Voltage Control Methods
This paper newly proposes the rapid voltage controls by the asynchronous commu-
nication by using the VISA standard [12]. The VISA is the standard for configuring,
programming, and troubleshooting instrumentation systems comprising GPIB, VXI,
PXI, Serial, Ethernet, and USB interfaces. Therefore, it helps the various programming
environments such as Microsoft Visual Studio or JDK (Java Development Kit) for the
multiple interfaces.
Figure 4 shows the program flow of the proposed asynchronous communication by
using the VISA standard. Although the previous prototype system [11] introduced the
synchronous communication for remoting the voltage machine, it was the problems to
adjust the beam-forming of the proposed AAA system. According to the previous study
indicated, it took more than one second to reach the voltage from zero volts to five
volts. If the vehicle run by 60 km/h, the vehicle will move about 16.7 m per one
second. Therefore, it was necessary to introduce more quick voltage remote system for
the prototype system.
In Fig. 4, first of all, the connection between the PC and the voltage machine is
established by the “OPEN” command of the VISA standard through the LAN. Then,
the remote control mode is activated by the “WriteString” command of the syn-
chronous communication, and the value of the target voltage is sent after receiving the
response from the voltage machine. Then, the current voltage setting is continuously
asking the voltage machine, and the connection is closed when the current voltage is
reached to the target voltage. Although the error functions are the future subjects for
this asynchronous controls, there were certain waiting times in case of the previous
controls, and the improvements of the proposed control methods are discussed in the
following chapters.
Fig. 4. The program flow of the proposed asynchronous communication by using the VISA
standard in order to control the beam-forming of the proposed AAA system.
4 Experiments
The experiments with the prototype system evaluate the effectiveness of the remote
voltage controls by the proposed asynchronous communications. In the experiments,
first of all, the processing time to reach the given voltage is observed by the system.
Table 1 shows the results of the processing time when the given voltage of the VISA
command is set to 1, 3, 5…15 V from 0 V, and the actual voltage output is shown in
the table.
Table 1. Processing time to reach the given voltage by the VISA commands.
Voltage changes Processing time (s) Observed voltage (V)
0 to 1 (V) 0.554 0.99
0 to 3 0.643 2.99
0 to 5 0.700 4.99
0 to 7 0.727 6.99
0 to 9 0.810 9.00
0 to 11 0.831 11.00
0 to 13 0.863 13.00
0 to 15 0.885 14.99
The previous paper [11] present that the processing time was 1.00 s from 0 V to
5 V, 2.70 s from 0 V to 5 V, and 3.69 s from 0 V to 15 V, but the proposed asyn-
chronous controls reached from 0 V to 15 V by 0.88 s in the experiments. According
to the results, it is effective that the proposed methods provide the quick beam-forming
controls in the prototype system, and it is suggested that the proposed methods are
useful for the V2V networks even if the vehicles moves so fast.
The reason for the rapid controls by the asynchronous command communication
could be the guaranteed time for waiting for the response from the voltage machine in
the VISA standards for the safety device controls, and so it is necessary to consider the
additional error controls in case of the malfunctions of the voltage machine or LAN
disconnections for the future studies.
Secondly, the processing time to control the given voltages and the angle of the
beam-forming by the prototype AAA system is shown in Fig. 5 and Table 2. Figure 5
indicates the differences between the processing time of the synchronous controls in the
previous study [11] and that of the proposed asynchronous controls.
Fig. 5. Results of the given voltages in the prototype AAA system in the experiments.
In the experiments, the voltage VISA commands were sent from the note PC to the
voltage machine by every 5.00 s, and the processing time of the increasing/decreasing
5.00 V was observed in the prototype system. The figure shows that the proposed
asynchronous controls rapidly reached to the given voltage at every point, and it is
supposed that the proposed methods are effective for the V2V communication. Also,
Table 2 shows the angle of the beam-forming in the AAA system and the processing
time in the experiments. It took only 0.855 s for changing 45° of the beam-forming in
the system, and it is also supposed that the proposed methods are more useful for the
rapid movements of the V2V communications.
Table 2. Processing time to reach the target voltage by the VISA commands.
Setting Voltage (V) Processing time (s) Peak angle of beam-forming (degree)
5 0.700 10.0
10 0.820 30.0
15 0.885 45.0
5 The Conclusion and Future Study
With the developments of ITS technologies such as V2V networks, it is expected to

realize the new kinds of applications such as the road surveillance system during the
winter. However, there are the limitations of the transmission range or the radio noise
in the use of the higher frequency for the V2V networks such as 5.9 GHz in the
IEEE802.11p standard. Therefore, this paper proposes the optimal AAA control
methods with the camera image recognition for the DTN method in the road surveil-
lance system in the winter, and it is especially discussed about the improvements of the
remote voltage control methods for the antenna elements.
This paper introduced the rapid voltage control methods for the Adaptive Array
Antenna with the asynchronous communication by using the VISA standard, and the
implementations of the proposed methods are explained and the experiments are held
for the effectiveness of the proposed methods. The results of the experiments show
great improvements in comparison with the previous researches, and it took less than
one second to change 45° of the beam-forming in the prototype system.
Now, the field experiments by the prototype system in the vehicles are planning in
the future studies, and the improvements of the vehicle recognition system in the
prototype system are also the future subjects.
Acknowledgement. This paper is the extend version of “Remote Voltage Controls by Image
Recognitions for Adaptive Array Antenna of Vehicular Delay Tolerant Networks” in the 13th
International Conference on Advances on P2P, Parallel, Grid, Cloud and Internet Computing
(3PGCIC2018), Oct. 2018. Also, this work was supported by SCOPE (Strategic Information and
Communications R&D Promotion Programme) Grant Number 181502003 by Ministry of
Internal Affairs and Communications in Japan.
References
1. Kato, S., Minobe, N., Tsugawa, S.: Experimental report of inter-vehicle communications
systems: a comparative and consideration of 5.8GHz DSRC, wireless LAN, cellular phone,
IEICE Techn. Rep. ITS 103(242), 23–28 (2012)
2. Uchida, N., Takahata, K., Shibata, Y., Shiratori, N.: Proposal of vehicle-to-vehicle based
delay tolerant networks with adaptive array antenna control systems. In: The 8th
International Workshop on Disaster and Emergency Information Network Systems
(IWDENS2016), pp. 649–654 (2016)
3. Welch, G., Bishop, G.: The Kalman Filter. http://www.cs.unc.edu/*welch/kalman/
4. Welch, G., Bishop, G.: An Introduction to the Kalman Filter. University of North Carolina at
Chapel Hill Department of Computer Science (2001)
5. Uchida, N., Ichimaru, R., Ito, K. Ishida, T., Shibata, Y.: Implementation of adaptive array
antenna controls with image recognitions for DTN based vehicle-to-vehicle networks. In:
The 9th International Workshop on Disaster and Emergency Information Network Systems
(IWDENS2017), pp. 633–638 (2017)
6. Aizawa, T., Shigeno, H., Yashiro, T., Matushita, Y.: A vehicle grouping method using W -
CDMA for road to vehicle and vehicle to vehicle communication. IEICE Technical
Report ITS 2000(83(2000-ITS-002)), pp. 67–72 (2000)
7. Karasawa, Y., Sekiguchi, T., Inoue, T.: The software antenna: a new concept of
Kaleidoscopic antenna in multimedia radio and mobile computing era. IEICE Trans.
Commun. E80-B, 1214–1217 (1997)
8. Yokoi, T., Iki, Y., Horikoshi, J., et al.: Software receiver technology and its applications.
IEICE Trans. Commun. E83-B(6), 1200–1209 (2000)
9. Fall, K., Hooke, A., Torgerson, L., Cerf, V., Durst, B., Scott, K.: Delay-tolerant networking:
an approach to interplanetary internet. IEEE Commun. Maga. 41(6), 128–136 (2003)
10. Uchida, N., Kawamura, N., Shibata, Y., Shiratori, N.: Proposal of Data Triage Methods for
Disaster Information Network System Based on Delay Tolerant Networking: The 7th
International Conference on Broadband and Wireless Computing, Communication and
Applications (BWCCA 2013), pp. 15–21 (2013)
11. Uchida, N., Kawamura, N., Shibata, Y., Shiratori, N.: Remote Voltage Controls by Image
Recognitions for Adaptive Array Antenna of Vehicular Delay Tolerant Networks: Advances
on P2P, Parallel, Grid, Cloud and Internet Computing. (3PGCIC 2018). Lecture Notes on
Data Engineering and Communications Technologies, vol. 24, pp. 387–394. Springer (2018)
12. National Instruments Coorporation: National Instruments VISA. https://www.ni.com/visa/
Collaborative Emerging Systems
(COLLABES-2019)
Design of Modern Logistics Management
System Based on RFID and NB-IoT
Jiayi Pang1 , Leixian Shen1 , Qingyun Zhang1 , He Xu1,2(B) , and Peng Li1,2
1
School of Computer, Nanjing University of Posts and Telecommunications,
Nanjing 210023, China
{b16041505,b16041522,b16041523,xuhe,lipeng}@njupt.edu.cn
2
Jiangsu High Technology Research Key Laboratory for Wireless Sensor Networks,
Nanjing 210003, China
Abstract. In view of the various drawbacks of the current logistics man-

agement methods using barcodes, this paper proposes a design scheme
of RFID-based logistics management system. Taking RFID as the core,
combining NB-IoT and deep learning technologies, this paper proposes a
smart logistics solution, which includes logistics tracking, auxiliary sort-
ing, fast loading and unloading, cargo positioning and inventory and
self-service pickup. This solution can improve the automation level of
the whole process of logistics. On the one hand, the highly automated
logistics link avoids manual operation errors and reduces labor costs.
On the other hand, using RFID tags instead of bar codes has laid the
foundation for fully automated storage and unmanned delivery.
1 Introduction
In the current express delivery industry, the barcode is still the main mean of
information storage and identification because of its simplicity and low cost, but
scanning barcodes requires a lot of manpower and material resources. Through-
out the logistics process, the barcode of each cargo needs to be scanned multiple
times. Moreover, because of the short recognition distance, only when the loca-
tion of the barcode is found can it be scanned, and only one can be scanned at
a time.
An express warehouse usually has a large throughput and fast circulation.
The traditional barcode has low efficiency of scanning and cannot support the
high-speed operation of the logistics industry. The RFID technology has the
advantages of non-contact, large data volume and long recognition distance, so
this paper proposes a logistics management system which replaces barcodes with
RFID tags.
2 Related Work
RFID can be applied in many aspects of logistics and will make great contribu-
tions to the sorting of cargoes and management of warehouses, which can greatly
improve the degree of automation and improve efficiency.
https://doi.org/10.1007/978-3-030-15035-8_54
562 J. Pang et al.
Cargo sorting is usually carried out on conveyor belts and requires the posi-
tioning of mobile RFID tags. There are many ways to identify and track RFID-
attached cargoes on conveyor belts. In [5], a new synthetic array phase-based
technique for localization and tracking of items equipped with UHF-RFID tags
are investigated. In [12], authors combining received signal strength indication
(RSSI) and Doppler frequency shift measurements to realize sorting for the con-
veyor belt scene of dense tags. It can be seen that the use of RFID RSSI and
other attributes to achieve cargo sorting is feasible and effective. In practical
applications, RFID-based airport bag sorting has been implemented.
When using RFID to manage warehouses, the core goal is to use RFID to
locate cargoes, thereby achieving forklift route optimization, handling warehouse
operation orders smartly, cargo inventory, etc. Cargo positioning is a kind of
indoor positioning, often using RSSI-based positioning algorithm [8–10], or using
phase-based algorithm [3,7,13], and in recent years there are also methods of
using machine learning [6,11]. NB-IoT has a lot of room for development as a
new technology. Due to its small bandwidth, long battery life and wide coverage,
it has great advantages in smart logistics [1,2].
3 Design of System
This section discusses the design of the whole system, including the system
architecture, data transmission and storage scheme and four logistics-related
schemes based on RFID.
3.1 System Architecture
The architecture of the entire logistics management system is shown in Fig. 1.

The system has seven layers, and Fig. 1 shows that each of them contains some
components and plays a different role in the system.
The system workflow is divided into two parts as described below. After
the user places the order, the staff enters the order information into the RFID
tag and attaches it to the cargo. Before and after logistics transportation, RFID
readers can be used for quick loading and unloading inspection, real-time position
monitoring of logistics vehicles. An RFID system is set up in the transfer station.
After cargoes enter the transfer station, the staff can use RFID technology to
quickly sorting the cargoes and use the indoor positioning scheme to achieve
quick inventory.
After many transfers, the cargoes entered the courier station. RFID readers
are placed around shelves, and an RFID-based smart outlet and camera are
placed at the door of the station. The work flow of the station is shown in Fig. 2.
First, the cargoes are put into storage, and the RFID-based quick loading and
unloading scheme can be used to check the errors. Then, after the staff put
the cargoes on the shelf, the system uses the indoor positioning technology to
realize automatic cargo positioning, and sends a pickup notification according to
the obtained cargo location information. Then, the consignee finds the cargoes
Design of Modern Logistics Management System 563
Fig. 1. System architecture diagram. The layers have different components and play
different roles and they are interconnected to form the entire system.
Fig. 2. Courier station workflow. A circle represents a person or an object, and a

rectangle represents an action or event.
564 J. Pang et al.
on the corresponding shelf according to the received short message, and scans
the cargoes on the self-service exit. Export’s camera takes a photo of a valid
document or face while scanning. Finally, when the staff needs to inventory the
cargoes, the RFID cargo positioning system will compare the positioning results
with the database information, and promptly find out the cargo that fails to be
scanned but the database shows that it has not been signed.
3.2 Data Transmission and Storage Scheme Based on NB-IoT
This scheme builds a narrow-band IoT network based on RFID reader, Rasp-
berry pie processor, NB-IoT module, operator platform, server, database. This
solution builds a logistics data transmission and storage solution, mainly based
on NB-IoT network, supplemented by WLAN and Ethernet networks.
3.2.1 Transmission and Storage of RFID Data

Many designs in this scheme are based on the reading of data in the RFID tag
by the RFID antenna, including readers of different frequencies. The data is
processed into the database through the middleware, and the flow of the RFID
data transmission and storage is shown in Fig. 3.
Fig. 3. Data transfer process.
3.2.2 Transmission and Storage of Logistics Vehicle Positioning Data

First, use the GPS/Beidou module to obtain the location information of the
logistics vehicle [4], including the vehicle number, latitude and longitude coor-
dinates, and then send the information to the background using the NB-IoT
module. After obtaining the data, the real-time location of the logistics vehicle
can be obtained by parsing the data to achieve the purpose of monitoring the
location of the logistics vehicle. The NB-IoT communication mode is shown in
Fig. 4.
Fig. 4. NB-IoT communication diagram.

3.3 RFID-Based Scheme
3.3.1 RFID-Based Cargo Sorting Scheme

For large automated warehouses, an RFID-based automatic sorting scheme can
be used. In other fields, there are many mature RFID-based automatic sorting
systems, airport baggage automatic sorting system, automatic book sorting sys-
tem, etc., which proves that RFID-based express automatic sorting scheme has
strong feasibility.
The automatic sorting process is as shown in Fig. 5. A plurality of sorting
boxes are arranged around the conveyor belt. After the cargoes are placed on the
conveyor belt, the antenna scans the RFID tags on the cargoes, and the EPC
code stored in the tags is transmitted to the background. The EPC code can
uniquely identify a certain item. The system background queries the detailed
information of the cargoes in the database according to the EPC code, and
determines the sorting box to which the cargoes should be sorted according to
the sorting code of the receiving address. The hardware facility automatically
completes the sorting when the cargoes arrive at the right location on the con-
veyor. For small warehouses, RFID assists manual sorting by scanning tags on
cargoes. Using the RFID antenna to scan the cargoes, the system will judge how
the cargoes should be sorted according to the RFID tag information, and inform
the sorter how to sort by text prompts, which can effectively improve the manual
sorting efficiency. Among them, the process of determining how the cargoes are
sorted is similar to automatic sorting. The difference is that the system displays
Fig. 5. Automatic sorting.

566 J. Pang et al.
the sorting result on the electronic screen, or calls the audio file to broadcast the
sorting result to the sorter, and finally sorts manually.
3.3.2 RFID-Based Loading and Unloading Scheme

Because RFID readers can read large quantities of RFID tags at the same time,
the system can read cargo information in batches and perform statistics to avoid
missed cargoes during loading and unloading. When cargoes check-in or check-
out, RFID antennas are used to scan cargoes, obtain the quantity and specific
information of the cargoes, compare the information of cargoes to be transported
in the database, and judge whether cargoes are missed or not, so as to solve the
problem in time.
3.3.3 RFID-Based Cargo Location and Inventory Scheme

At present, there are many indoor positioning methods using phase or received
signal strength, as well as TOA, AOA and other positioning mechanisms. The
positioning scheme proposed in this paper employs deep learning method, using
phase, received signal strength and other data, utilities reference tags to position
cargoes relatively and ultimately it can identify the layer where a cargo is located.
As shown in Fig. 6, in the experimental stage, the two-dimensional plane is
used to simulate the shelf, and the position of the cargo tag is changed multiple
times. The phase and the received signal strength of the tag of the different
position points are read. Then the data will be processed. The phase and the
received signal strength of each cargo tag respectively subtract those of the refer-
ence tag, so that the relationship data between each cargo tag and the reference
tag is obtained. The relationship data of each cargo tag and the corresponding
shelf number are input into the deep neural network for training, and finally,
a general model of the relationship between the cargo location and the refer-
ence tag is output. In the plane of the experiment, the system can accurately
position cargoes, determine which shelf cargoes are on, and visually display the
positioning results.
Fig. 6. Cargo positioning simulation. Fig. 7. Cargo positioning scenario.

Several reference tags are put on the sim-
ulated shelf, and one tag is the cargo to
be positioned.
The actual cargo positioning scenario is shown in Fig. 7. After obtaining the
positioning information of a cargo tag, the information of each cargo tag read
out is compared with the reference tag information, and the position on the shelf
can be judged according to the model. Finally, Update the location information
of cargoes in the database.
After the staff put cargoes on the shelf, the RFID antenna collects the tag
information on cargoes and transmits them to the cargo positioning system. The
cargo positioning system uses the algorithm to obtain the shelf where cargoes
are located, and automatically sends a pickup notification message according to
the recipient information stored in the tag information, and the consignees can
self-service receive cargoes. At the same time of locating cargoes, the system
performs the express delivery at a certain frequency. For the express delivery
that has not been scanned and has not been signed, the staff is immediately
reminded to check whether cargoes are lost.
3.3.4 RFID-Based Self-service Cargo Receiving Scheme

The smart exit includes an RFID reader and a camera connected to the Rasp-
berry Pi. The receiver picks up the courier and then goes to the smart exit. The
exported RFID reader scans RFID tags of cargoes, and the camera automatically
captures the document or face and records the information of the receiver. The
process is shown in Fig. 8. Compared with manual operation, self-service cargo
receiving records the pick-up information by shooting documents or face photos,
simplifies the pick-up process, improves the efficiency of pick-up, and solves the
problem that the receiver does not carry the documents when receiving cargoes.
Fig. 8. Self-service cargo receiving process.
4 Experiment Results and Performance Analysis
4.1 NB-IoT Platform Sends and Receives Data

NB-IoT platform receiving and transmitting vehicle positioning experimental
data is shown in Fig. 9. In the experiment, the platform can successfully accept
and send positioning information, the information is mainly latitude and longi-
tude which is used for vehicle positioning described in Sect. 3.2.2
568 J. Pang et al.
Fig. 9. NB-IoT platform sends and receives data.
4.2 RFID-Based Cargo Location

RFID-based cargo location scheme is described in Sect. 3.3.3. The experimental
to-be-positioned area shown in Fig. 6 is 40 cm × 60 cm, and the final stable
accuracy of the model is about 95%. Besides, the accuracy of other sizes of
positioning areas is shown in Table 1. As can be seen from the results, the cargo
location scheme achieves good accuracy and has the possibility of being applied
to the actual scene.
Table 1. Experimental results for different areas to be located
Size of the area to be located Number of Number of Accuracy

test cases correct cases
40 cm × 60 cm 60 57 95%
30 cm × 40 cm 50 47 94%
20 cm × 30 cm 40 36 90%
15 cm × 20 cm 40 34 85%
12 cm × 12 cm 40 32 80%
6 cm × 6 cm 40 31 78%
5 Conclusions
First of all, the modern logistics management system proposed in this paper
uses the express mail with RFID tags, which greatly improves the automation
of the whole process of logistics. Secondly, the system uses NB-IoT technol-
ogy as the end communication method, which can reduce the communication
cost, and is also the development and expansion of NB-IoT technology. Finally,
replacing barcodes with RFID tags can lay the foundation for fully automated
warehousing and unattended distribution. Under the background of the Internet
of Things, the logistics management system proposed in this paper is in line with
the inevitable trend of smart logistics development, and truly realizes the goal
of serving the society with science and technology.
Acknowledgements. This work is financially supported by the National Natural

Science Foundation of P. R. China (Nos. 61572260, 61572261, 61672296, 61602261),
CERNET Innovation Project (No. NGII20180605), Scientific & Technological Support
Project of Jiangsu Province (Nos. BE2015702, BE2016185, BE2016777), Postgraduate
Research and Practice Innovation Program of Jiangsu Province (No. KYCX 17 0798,
No. KYCX18 0931) and NUPT STITP (No. SZDG2018014).
References
1. Adhikary, A., Lin, X., Wang, Y.P.E.: Performance evaluation of NB-IoT coverage.
In: 2016 IEEE 84th Vehicular Technology Conference (VTC-Fall), pp. 1–5. IEEE
(2016)
2. Chen, J., Hu, K., Wang, Q., Sun, Y., Shi, Z., He, S.: Narrowband internet of things:
implementations and applications. IEEE Internet Things J. 4(6), 2309–2314 (2017)
3. Hekimian-Williams, C., Grant, B., Liu, X., Zhang, Z., Kumar, P.: Accurate localiza-
tion of RFID tags using phase difference. In: 2010 IEEE International Conference
on RFID, pp. 89–96. IEEE (2010)
4. Huai, Y., Shao, Q., Lu, Z.: Application research of UM220 Beidou/GPS hybrid
position module. Foreign Electron. Meas. Technol. 33(03), 76–79 (2014)
5. Nepa, P., Lombardini, F., Buffi, A.: Location and tracking of items moving on a
conveyor belt and equipped with UHF-RFID tags. In: 2012 IEEE Antennas and
Propagation Society International Symposium (APSURSI), pp. 1–2. IEEE (2012)
6. Ng, W.W., Lin, L., Chan, P.P., Yeung, D.S.: 3D goods allocation in warehouse
with L-GEM based 3-D RFID positioning. In: 2011 International Conference on
Machine Learning and Cybernetics (ICMLC), vol. 1, pp. 324–329. IEEE (2011)
7. Shangguan, L., Yang, Z., Liu, A.X., Zhou, Z., Liu, Y.: Relative localization of RFID
tags using spatial-temporal phase profiling. In: NSDI, pp. 251–263 (2015)
8. Siachalou, S., Bletsas, A., Sahalos, J., Dimitriou, A.G.: RSSI-based maximum like-
lihood localization of passive RFID tags using a mobile cart. In: 2016 IEEE Wireless
Power Transfer Conference (WPTC), pp. 1–4. IEEE (2016)
9. Subedi, S., Pauls, E., Zhang, Y.D.: Accurate localization and tracking of a passive
RFID reader based on RSSI measurements. IEEE J. Radio Freq. Identif. 1(2),
144–154 (2017)
10. Wang, G., Qian, C., Shangguan, L., Ding, H., Han, J., Yang, N., Xi, W., Zhao,
J.: HMRL: relative localization of RFID tags with static devices. In: 2017 14th
Annual IEEE International Conference on Sensing, Communication, and Network-
ing (SECON), pp. 1–9. IEEE (2017)
11. Xu, H., Ding, Y., Li, P., Wang, R.: Indoor localization using the reference tags and
phase of passive UHF-RFID tags. Int. J. Bus. Data Commun. Network. (IJBDCN)
13(2), 69–82 (2017)
12. Xu, J., Yang, L., Liu, Q., Hu, J., Song, T.: A sorting algorithm for RFID tags mov-
ing on a conveyor belt. In: 2018 IEEE International Conference on RFID (RFID),
pp. 1–7. IEEE (2018)
13. Yang, L., Chen, Y., Li, X.Y., Xiao, C., Li, M., Liu, Y.: Tagoram: real-time tracking
of mobile RFID tags to high precision using cots devices. In: Proceedings of the
20th Annual International Conference on Mobile Computing and Networking, pp.
237–248. ACM (2014)
Evaluation of TBC and OBC Precedent
Relations Among Messages in P2P Type
of Topic-Based Publish/Subscribe System
Takumi Saito1(B) , Shigenari Nakamura1 , Dilawaer Duolikun1 ,

Tomoya Enokido2 , and Makoto Takizawa1
1
takumi.saito.3j@stu.hosei.ac.jp, nakamura.shigenari@gmail.com,
dilewerdolkun@gmail.com, makoto.takizawa@computer.org
2
eno@ris.ac.jp
Abstract. In a P2PPS (peer-to-peer model of a topic-based pub-

lish/subscribe) system, peer processes (peers) are cooperating by
exchanging messages with no centralized coordinator. Each peer pub-
lishes a message carrying objects. Messages which have a common object
are related. After receiving messages, a peer has to causally deliver
related messages. Contents of messages are denoted by publication top-
ics. Each peer specifies subscription topics and receives messages whose
publication topics include some subscription topic. A pair of messages
which have a common publication topic are related. Then, we define a
topic-based-causally (TBC) precedent relation among messages which
are related in terms of topics. Then, we newly propose a TBC protocol
to deliver messages to target peers by using topic vector and linear time.
However, even if m1 is delivered before m2 in the TBC protocol, m1
may not TBC-precede m2 . Here, a pair of the messages m1 and m2 are
unnecessarily ordered in a common target peer. In the evaluation, we
show number of messages unnecessarily ordered in the TBC protocol.
Keywords: Topic-based publish/subscribe system · P2P model ·

OBC-precedent relation · TBC-precedent relation · TBC-protocol ·
Unnecessarily ordered messages
1 Introduction
In a publish/subscribe (PS) model [2,6,18,19], processes are cooperating by

exchanging messages with one another. In this paper, we consider a peer-to-peer
(P2P) model of a topic-based PS (P2PPS) system [10–12,17] where there is no
centralized coordinator. Let T be a set of topics in a system. Each peer pi sub-
scribes a collection pi .ST (⊆ T ) of subscription topics. Each peer pi publishes
a message m with publication topics m.P T (⊆ T ) which denote contents of the

https://doi.org/10.1007/978-3-030-15035-8_55
A Causally Precedent Relation 571
message. A message m is only received by a target peer pi where the publica-

tion m.P T and the subscription pi .ST include some common topic. Messages
published by a peer are received by every target peer in the publishing order
by underlying networks like TCP [5]. However, messages published by different
peers may not be received by different target peers in the same order.
An object is a unit of data resource in a system. A peer pi publishes a message
which carries objects to target peers. The meanings of objects are denoted by
topics. The publication topics m.P T of a message m are a collection of topics
of objects carried by the message m. Each peer pi receives a message m where
m.P includes some subscription topic in pi .ST . A pair of messages m1 and m2
which carry a common object are related. Suppose a peer pi receives a message
m1 and then publishes a message m2 . Here, the message m1 causally precedes
the message m2 (m1 → m2 ) according to the traditional causality theory [8,14].
However, if a pair of messages m1 and m2 are not related, the causally precedent
relation m1 → m2 is meaningless. Here, a message m1 object-based-causally
(OBC) precedes a message m2 (m1 ⇒ m2 ) iff m1 and m2 are related and m1
→ m2 . Each target peer is required to deliver a pair of received messages in the
OBC-precedent order.
In order to OBC-deliver messages [15,16], each peer pi has to keep objects
of messages received in a memory and compare the objects with objects in the
memory. It is not easy for each peer to hold and manipulate every object received.
Objects are characterized by topics and the publication topics of messages are
topics of objects in the messages. Based on the OBC-precedent relation, we
define a topic-based-causally (TBC) precedent relation among messages in terms
of topics. That is, a message m1 TBC-precedes a message m2 (m1 m2 ) iff (if
and only if) m1 → m2 and m1 .P T ∩ m2 .P T = φ. A message m1 TBC-precedes
a message m2 if m1 ⇒ m2 . If a pair of messages m1 and m2 carry a common
object o1 , the messages m1 and m2 have common publication topics, i.e. topics
of objects. However, if m1 and m2 carry different objects o1 and o2 , respectively,
which are denoted by a same topic t, the messages m1 and m2 are related in
terms of topics. However, the messages m1 and m2 are not related in terms of
objects. Hence, even if m1 TBC-precedes m2 , m1 may not OBC-precede m2 .
We propose a TBC protocol where messages received are ordered by using the
linear clock [8] and topic vector [13]. A message m1 precedes a message m2 in
the TBC protocol (m1 ≺ m2 ) if m1 TBC-precedes m2 in reality. However, even
if m1 ≺ m2 in the TBC protocol, m1 may not TBC-precede m2 . Messages m1
and m2 are unnecessarily ordered in a peer pi if pi receives m1 before m2 and
m1 ≺ m2 , but m1 does not TBC-precede m2 . In the evaluation, we show the
number of pairs of messages unnecessarily ordered in the TBC protocol.
In Sect. 2, we present a system model. In Sect. 3, we discuss the OBC-
precedent relation among messages. In Sect. 4, we propose the TBC protocol.
In Sect. 5, we evaluate the TBC protocol.
572 T. Saito et al.
2 System Model
In P2PPS (peer-to-peer publish/subscribe) systems [1,11–13,17], peer processes
(peers) p1 , ..., pn (n ≥ 1) are cooperating with one another by exchanging mes-
sages in underlying networks. Here, there is no centralized coordinator [3] to
do the communication among peers. Each peer pi can publish messages while
receiving messages from other peers. A message m is only received by a target
peer which is interested in the contents of the message m.
In this paper, peers exchange objects with one another by publishing and
receiving messages. Contents [4,7] of a message m are considered to be a collec-
tion of objects carried by the message m. Let O be a set of objects in a system.
An object is a unit of data resource. Let m.O (⊆ O) be a set of objects carried
by a message m. A peer which is interested in the contents m.O of a message
m is a target peer of the message m. A peer pi supports a memory pi .M to
store objects. A peer pi creates and manipulates objects in the memory pi .M .
On receipt of a message m, a peer pi stores objects m.O carried by the message
m in the memory pi .M . In this paper, a pair of object collections O1 and O2 are
related if O1 ∩ O2 = φ, i.e. O1 and O2 include a common subobject. A transac-
tion is a unit of work of each peer pi . A transaction is a sequence of publication
and receiving events of messages. A peer pi is modeled to be a sequence of trans-
actions Ti1 , ..., Tili . This means, objects are not persistent. Messages published
and received in a pair of different transactions Tik and Tih are not related. That
is, each time a transaction starts on a peer pi , the memory pi .M is flushed.
We assume the underlying network supports a pair of peers pi and pj with
the following communication service:
1. Each peer pi receives only and all messages in whose contents the peer pi is
interested. Here, the peer pi is a target peer of the message m.
2. Messages published by each peer can be received by every target peer in the
publishing order without message loss and duplication.
3. Every message published by a peer arrived at every peer.
If a peer pi publishes a message m1 before another message m2 , every com-
mon target peer pj receives the messages m1 then m2 in the publishing order.
However, some pair of messages m1 and m2 published by different peers may not
be received by every common destination peer in the same order. For example,
a peer pi publishes a message mi1 then mi2 and another peer pj publishes a
message mj1 then mj2 . Suppose a pair of peers pk and pl are common target
peers of the messages mi1 , mi2 , mj1 , and mj2 . Here, the target peer pk receives
the messages mi1 , mj1 , mj2 , and mi2 in the sequence. On the other hand, the
other peer pl receives the messages in a sequence mj1 , mi1 , mj2 , and mi2 . Here,
each peer receives a pair of messages mi1 and mi2 from the peer pi and the
messages mj1 and mj2 from other the peer pj in the publishing order. However,
the peer pk receives mi1 before mj1 but the peer pl receives mj1 before mi1 . If
the message mi1 causally precedes the message mj1 (mi1 → mj1 ), the peer pl
has to deliver mi1 before mj1 as the peer pk . Each peer has to causally deliver
only and every pair of related messages.
Each object oi is characterized in terms of topics. Let oi .T (⊆ T ) be a subset

of topics of an object oi . A peer pi subscribes interesting topics. pi .ST (⊆ T ) is
a set of subscription topics of the peer pi . The publication m.P T of a message
m is a set of topics of objects in the message m, i.e. m.P T = ∪oi ∈m.O oi .T . A
peer pi receives a message m only if the message m carries topics subscribed by
the peer pi , i.e. m.P T ∩ pi .ST = φ. Here, pi is a target peer of the message m.
A peer pi publishes a message m after including objects in the memory pi .M .
Let us consider a pair of messages m1 and m2 . If m1 .O and m2 .O include a
common object o, m1 .P T and m2 .P T include common topics o.T of the common
object o, i.e. m1 .P T ∩ m2 .P T ⊇ o.T (= φ). Next, suppose the messages m1 and
m2 only include objects o1 and o2 , respectively, and o1 .T ∩ o2 .T = φ. Here,
m1 .P T (= o1 .T ) ∩ m2 .P T (= o2 .T ) = φ but the messages m1 and m2 are not
related in terms of objects since m1 .O ∩ m2 .O = φ. Thus, if a pair of messages
m1 and m2 are related, m1 .P T ∩ m2 .P T = φ. However, even if m1 .P T ∩ m2 .P T
= φ, m1 and m2 may not be related.
3 An Object-Based-Causally (OBC) Precedent Relation

of Messages
If a peer pi publishes a message mi after receiving another message mj published
by a peer pj , the message mi causally precedes the message mj (mi → mj )
since the publication event of the message mi happens before the message mj
according to the causality theory [8]. Here, the message mj carries a set mj .O of
objects to the target peer pi . On receipt of the message mj , the peer pi stores the
objects mj .O in the memory pi .M . After receiving the message mj , the target
peer pi manipulates objects in the memory pi .M . Then, the peer pi publishes
a message mi including objects in the memory pi .M . Suppose an object oj (∈
mj .O) carried by a message mj is also carried by another message mi . Here, the
message mi is related with the message mj since oj ∈ mi .O ∩ mj .O. If no object
in the message mj carried by the message mj , the message mi is not related
with the message mj . Thus, a message mi is related with a message mj iff mi .O
∩ mj .O = φ.
Next, we define an object-based-causally (OBC) precedent relation ⇒
among messages.
[Definition] Let mi and mj be a pair of messages and A be a subset of objects
(A ⊆ mi .O ∩ m.j.O).
1. mi primarily object-based-causally (OBC) precedes mj (mi →A mj ) with

respect to A iff (if and only if) a peer pi publishes the message mj after
receiving the message mi .
2. mi object-based-causally (OBC) precedes mj (mi ⇒A mj ) with respect to
A iff one of the following conditions holds:
• mi primarily OBC-precedes mj with respect to the subset A (mi →A
mj ).
• mi →A mk and mk ⇒A mj for some message mk .
574 T. Saito et al.
Fig. 1. OBC-precedency.
3. mi and mj are OBC-concurrent with respect to A (mi A mj ) iff nether mi

⇒A mj nor mj ⇒A mi .
A message mi primarily OBC-precedes a message mj (mi → mj ) iff mi →A

mj for some subset A (⊆ mi .O ∩ mj .O) of objects. A message mi OBC-precedes
a message mj (mi ⇒ mj ) iff mi ⇒A mj for some subset A (⊆ mi .O ∩ mj .O) of
objects. A pair of messages mi and mj are OBC-concurrent (mi mj ) iff mi
A mj for some subset A of objects.
Figure 1 shows four peers pi , pj , pk , and ph which are cooperating with one
another. The peer first pi publishes a message mi with a pair of objects o1 and
o2 , i.e. mi .O = {o1 , o2 }. Then, the peer pj publishes a message mj where mj .O
= {o2 , o3 }. Here, the message mi causally precedes the message mj (mi →
mj ). Since o2 ∈ mi .O ∩ mj .O and the message mi carries the object o2 in the
message mj , the message mi OBC-precedes the message mj with respect to o2
(mi ⇒{o2 } mj ).
From the definitions, m1 ⇒A m2 if m1 → m2 . It is not easy for each peer pi to
keep in record every object carried by messages. It also takes time for each peer pi
to check if objects of each message are related with objects in the memory pi .M .
We consider a topic-based PS system [17] where each peer pi receives messages
only based on publication topics of the messages and the subscription topics of
the peer pi on behalf of objects. It is simpler to decide on which message causally
precedes another message by using publication topics carried by the messages
and subscription topics of peers.
4 A TBC (Topic-Based Causally Ordering) Protocol
4.1 A TBC Precedent Relation
In a topic-based P2PPS system, a message m with publication topics m.P T is

received by a target peer pi if m.P T ∩ pi .ST = φ. Suppose a pair of peers pi
and pj publish messages mi and mj , respectively, and each of target peers ph
Fig. 2. OBC-precedent relation on messages.
and pk receives both the messages mi and mj . Here, the peer ph first receives
the message mi and then the message mj . On the other hand, another peer pk
first receives the message mj and then the message mi in an order different from
the peer ph . Thus, messages published by different peers may be received by
different common target peers in different orders depending on network delay
time because there is no centralized coordinator. If the message mi OBC-precedes
the message mj (mi ⇒ mj ), each common target peer pk is required to receive
the message mi before the message mj . However, it is not easy for each peer
to hold every object received and manipulate objects to check if the objects are
related. In this paper, we newly define a topic-based-causally (T BC) precedent
relation.
[Definition] Let m1 and m2 be a pair of messages and B be a subset of topics
(B ⊆ m1 .P T ∩ m2 .P T ).
1. m1 primarily T BC-precedes m2 with respect to B (m1 →B m2 ) iff m1 is
received and then m2 is published before m2 .
2. m1 T BC-precedes m2 with respect to B (m1 B m2 ) iff one of the following
conditions holds:
• m1 →B m2 .
• m1 →B m3 and m3 B m2 for some message m3 .
3. m1 is TBC-concurrent with m2 with respect to B (m1 |||B m2 ) iff nether m1
B m2 nor m2 B m1 .
Suppose m1 →B1 m2 →B2 ... →Bk−1 mk . Here, it is noted m1 B mk may
not hold. m1 B mk means that the topic sets B1 , ..., Bk−1 include common
topics B, i.e. B ⊆ B1 ∩ ... ∩ Bk−1 .
A message m1 TBC-precedes a message m2 (m1 m2 ) iff m1 B m2
for some subset B (⊆ T ) of topics. A pair of messages m1 and m2 are TBC-
concurrent (m1 ||| m2 ) iff m1 |||B m2 for some subset B (⊆ T ).
576 T. Saito et al.
Suppose there are four peers pi , pj , ph , and pk as shown in Fig. 3. The peer pi
first publishes a message mi . The peer pj publishes a message mj after receiving
the message mi . Lastly, the peer pk publishes a message mk after receiving the
message mj . Suppose mi .P T = {a, b}, mj .P T = {a, c}, and mk .P T = {a, d}.
Here, the message mi primarily TBC-precedes the message mj (mi →{a} mj )
since mi → mj and {a} ⊆ mi .P T ∩ mj .P T . Hence, the message mi TBC-
precedes a pair of the messages mj and mk with respect to a topic a, i.e. mi
{a} mj and mi {a} mk . It is noted, each of the messages mi , mj , and mk
includes the publication topic a and mi →{a} mj and mj →{a} mk .
Fig. 3. TBC-precedent relation.
The message mi OBC-precedes the message mj (mi ⇒ mj ) and a pair of the

messages mi and mj carry the object o2 as shown in Fig. 2. Here, the objects o1
and o2 carried by the message mi are denoted by topics o1 .T = {a} and o2 .T =
{b}, respectively, and the object o3 carried by the message mj is denoted by a
topic o3 .T = {c}. Here, mi .P T = {a, b} and mj .P T = {b, c}. As shown in Fig. 4,
the message mi primarily TBC-precedes the message mj with respect to the
topic b (mi →{b} mj ) since the peer pi publishes the message mj after receiving
the message mi and mi .P T ∩ mj .P T = {b}. Thus, if mi ⇒{o2 } mj , mi →{b} mj .
Next, suppose the message mj carries a pair of the objects o3 and o4 where
o4 .T = {b}. Hence, mj .P T = {b, c}. However, the messages mi and mj carry
the objects o2 and o4 of the topic b, respectively, but the objects o2 and o4 are
different. This means, the messages mi and mj are not related. Thus, mi →{b}
mj but mi mj .
In Fig. 4, suppose a pair of messages mi and mj carry objects, i.e. mi .O =
{o1 } and mj .O = {o4 }. Since, o1 .T = {a, b}, and o4 .T = {b, c}, mi .P T = {a, b},
mj .P T = {a, b}. Although mi .P T ∩ mj .P T = {b}, mi .O ∩ mj .O = φ. Thus,
even if mi {b} mj , the OBC-precedent relation (mi ⇒ mj ) does not hold.
Fig. 4. TBC-precedent relation.
4.2 TBC Protocol
We newly propose a TBC protocol to TBC-deliver messages in a P2PPS system

composed of n (≥ 1) peers p1 , ..., pn . We assume every message is reliably
broadcast to every peer pi in a system. That is, every message published by a
peer arrives at every peer in a system in the publishing order without message
loss and duplication.
In the TBC protocol, each message m arrives at every peer in an underlying
network. Then, a peer pi receives a messages m if m.P T ∩ p.ST = φ. Publication
m.P T and subscription pi .ST are realized in a bitmap b1 , ..., bl where each bit
bj is 0 or 1. If a topic tk is included in a publication a subscription, the kth bit
bk is 1 in the bitmap, otherwise bk = 0.
In papers [10–13], the topic vector is proposed to TBC-deliver messages.
There are topics t1 , ..., tl (l ≥ 1), i.e. T = {t1 , ..., tl } in a system. Each peer pi
manipulates a topic vector T V = tv1 , ..., tvl for a set T of the topics t1 , ..., tl .
Each element tvk is manipulated for a topic tk . Initially, each element tvk is 0.
For a pair of topic vectors A = a1 , ..., al and B = b1 , ..., bl and a subset
V (⊆ T ) of the topics;
1. A >V B iff ak > bk for every topic tk in V .

2. A ≡V B iff ak = bk for every topic tk in V .
3. A ≥V B iff A >V B or A ≡V B.
4. A |V B (A and B are uncomparable) iff neither A ≥V B nor A ≤V B.
Suppose a peer pi would like to publish a message m. An element tvk is incre-

mented by one for every topic tk in m.P T . The message m carries the topic
vector in a field m.T V , i.e. m.T V = T V . Then the message m is published. On
arrival of a message m, tvk = max(tvk , m.tvk ) for every topic tk in every peer pi .
In this paper, messages are causally ordered in the linear clock [8] since the
size of the linear time is constant, i.e. independent of the numbers of peers and
topics. Each peer pi manipulates a variable L showing linear time. Each message
578 T. Saito et al.
m carries the linear time m.L when a peer sends the message m. On receipt of
a message m, the liner time L of a peer pi is max (L, m.L). In order to publish
a message m, a peer pi increments the variable L by one, i.e. L = L + 1, and
m.L = L. Then, the peer pi publishes the message m. Suppose a peer pi receives
a pair of messages m1 and m2 . If m1 .L < m2 .L, the message m1 is delivered
before the message m2 in the target peer pi . Otherwise, the messages m1 and
m2 are delivered in the receipt order.
Suppose a peer pi publishes a message m with publication m.P T . A message
m carries a topic vector m.T V and linear time m.L. The message m is published
by the peer pi as follows:
[Publication at a peer pi ]
1. m.O = objects which m carries (⊆ pi .M );

2. m.P T = topics of objects in m.O;
3. tvk = tvk + 1 for every topic tk in the publication m.P T ;
4. m.T V = T V ;
5. L = L +1; m.L = L;
6. publish m;
On arrival of a message m, a peer pi manipulates the topic vector T V to receive

the message m as follows:
[Receipt at a peer pi ]
1. A message m arrives at a peer pi .

2. tvk = max (tvk , m.tvk ) for every topic tk ;
3. L = max (L, m.L);
4. receive m in a receipt queue RQ if m.T ∩ pi .ST = φ, else neglect m;
5. Objects in m.O are stored in the memory pi .M .
For a pair of messages m1 and m2 received by a peer pi , m1 precedes m2 (m1

≺ m2 ) in the peer pi iff m1 .L < m2 .L and m1 .T V ≤A m2 .T V for topics A =
m1 .P T ∩ m2 .P T . In the receipt queue RQ, message are ordered in the precedent
relation ≺.
A message m1 precedes a message m2 in the TBC protocol (m1 ≺ m2 ) if
m1 B m2 for topics B = m1 .P T ∩ m2 .P T . This means, even if m1 ≺ m2 ,
“m1 A m2 ” may not hold. Suppose a peer pi receives a message m2 before
m1 . If m1 ≺ m2 , the peer pi cannot deliver the message m2 until receiving the
message m1 . Unless m1 B m2 for B = m1 .P T ∩ m2 .P T , there is no need for
the peer pi to wait for the message m1 . Here, a pair of the messages m1 and m2
are unnecessarily ordered in the peer pi .
[Definition] A pair of messages m1 and m2 are unnecessarily ordered with
respect to topics by a target peer pi in the TBC protocol iff m2 is received
before m1 and m1 ≺ m2 in the peer pi but m1 B m2 for topics B = m1 .P T
∩ m2 .P T .
We have to reduce the number of messages unnecessarily ordered. The more
number of messages are unnecessarily ordered, the larger messages are delayed
to be delivered.
5 Evaluation
We evaluate the TBC protocol in terms of number of pairs of messages unnec-
essarily ordered with respect to topics. A system includes peers p1 , ..., pn
(n ≥ 1). Let T be a set of topics t1 , ... tl (l ≥ 1). Let O be a set of objects
o1 , ... ol (l ≥ 1).
Each peer pi randomly publishes messages. That is, the publishing time
m.P BT of a message m published by a peer pi is randomly decided from time 0
to time maxT [time unit (tu)]. In the evaluation, value of maxT is 1000. At each
time unit, a peer pi is randomly decided in all the peers. Then, the peer pi cre-
ates one message m and publishes the message m. The receiving time m.RV Tj
of a message m of each peer pj is decided as m.RV Tj = m.P BT + δij . Here,
the delay time δij between a pair of peers pi and pj is randomly decided from
1 to 10 [tu]. Each message m carries the topic vector m.T V in addition to the
linear time m.L for evaluation.
In the evaluation, publication and subscription topics of each peer pi are
randomly taken from the topic set T . Each peer pi creates only one object o
where the publication o.T is a subset of subscription topics of the peer pi . Then,
the peer pi creates an object o in the set O with probability 0.01. Publication
topics of a message m are topics of objects m.O. In the evaluation, each peer
pi manipulates the vector time variable V T [9] and each message m carries the
vector time m.V T to check the causally precedent relation → among messages.
Each peer pi includes the vector time V T in a message m as m.V T . In the
vector time, “mi .V T < mj .V T ” iff a message mi causally precedes a message
mj (mi → mj ). Even if mi .LT < mj .LT holds, mi → mj may not hold. Here,
if a message mj is received before another message mi , a pair of the messages
mi and mj are unnecessarily ordered.
Suppose a peer pi receives a message m1 before a message m2 but m2 ≺
m1 . In the TBC protocol, the messages m1 and m2 are exchanged in the receipt
queue. Here, if m2 .V T < m1 .V T does not hold, the messages m1 and m2 are
unnecessarily ordered. In the simulation, we obtain the number of pairs of mes-
sages unnecessarily ordered.
Figure 5 shows the number of pairs of messages unnecessarily ordered. Here,
there are 10 peers (n = 10) and 20 topics (l = 20). Subscription of each peer is
randomly selected 1 to 5 topics from all topics (mstn = 5). The number of pairs
of messages unnecessarily ordered linearly increases for the number of messages
published by peers.
580 T. Saito et al.
Fig. 5. Evaluation.
In this paper, we discussed the P2PPS (peer-to-peer model of the topic-based

publish/subscribe) system. First, we defined the causally precedent relation on
messages in terms of objects carried by messages. A message m1 OBC-precedes
a message m2 (m1 ⇒ m2 ) iff the message m1 causally precedes the message m2
and the messages m1 and m2 carry a common object. Then, we defined the TBC-
precedent relation among messages in terms of topics. We proposed the TBC
protocol where messages are ordered by using the linear time and topic vector. A
message m1 precedes a message m2 (m1 ≺ m2 ) in the TBC protocol if m1 TBC-
precedes m2 (m1 m2 ). However, even if m1 ≺ m2 , m1 m2 may not hold. In
the evaluation, we showed the number of pairs of messages unnecessarily ordered
in the TBC protocol. We are now discussing a pair of messages unnecessarily
ordered in the TBC protocol with respect to the OBC-precedent relation.

15H0295.
References
1. Google alert. http://www.google.com/alerts
2. Google cloud pub/sub. https://cloud.google.com/pubsub/docs/overview?hl=ja
3. Arridha, R., Sukaridoho, S., Pramadihanto, D., Funabiki, N.: Classfiation IoT-big
data analytic for smart environment monitoring and analytic in real-time system.
Int. J. Space-Based Situated Comput. (IJSSC) 7(2), 82–93 (2017)
4. Bouramoul, A.: Contextualisation of information retrieval process and document
ranking task in web search tools. Int. J. Space-Based Situated Comput. (IJSSC)
6(2), 74–89 (2016)
5. Comer, D.E.: Internetworking with TCP/IP (VOLUME I). Prentice Hall, Upper
Saddle River (1991)
6. Eugster, P., Felber, P., Guerraoui, R., Kermarrec, A.: The many faces of pub-
lish/subscribe. ACM Comput. Surv. 35(2), 114–131 (2003)
7. Jacobson, V., Smetters, D., Thornton, J., Plass, M., Briggs, N., Braynard, R.:
Networking named content. In: Proceedings of the 5th International Conference
on Emerging Networking Experiments and Technologies (CoNEXT 2009), pp. 1–
12 (2009)
8. Lamport, L.: Time, clocks, and the ordering of event in a distributed systems.
Commun. ACM 21(7), 558–565 (1978)
9. Mattern, F.: Virtual time and global states of distributed systems. In: Parallel and
Distributed Algorithms, pp. 215–226 (1988)
10. Nakamura, S., Enokido, T., Takizawa, M., Ogiela, L.: An information flow control
models in a topic-based publish/subscribe systems. J. High Speed Netw. 24(3),
243–257 (2018)
11. Nakayama, H., Duolikun, D., Enokido, T., Takizawa, M.: Selective delivery of event
messages in peer-to-peer topic-based publish/subscribe systems. In: Proceedings of
the 18th International Conference on Network-Based Information Systems (NBiS-
2015), pp. 379–386 (2015)
12. Nakayama, H., Duolikun, D., Enokido, T., Takizawa, M.: Reduction of unnecessar-
ily ordered event messages in peer-to-peer model of topic-based publish/subscribe
systems. In: Proceedings of IEEE the 30th International Conference on Advanced
Information Networking and Applications (AINA-2016), pp. 1160–1167 (2016)
13. Nakayama, H., Ogawa, E., Nakamura, S., Enokido, T., Takizawa, M.: Topic-based
selective delivery of event messages in peer-to-peer model of publish/subscribe sys-
tems in heterogeneous networks. In: Proceedings of the 18th International Confer-
ence on Network-Based Information Systems (WAINA-2017), pp. 1162–1168 (2017)
14. Raynal, M.: Distributed Algorithms for Message-Passing Systems. Springer,
Heidelberg (2013)
15. Saito, T., Nakamura, S., Enokido, D.D.T., Takizawa, M.: Object-based selective
delivery of event messages in topic-based publish/subscribe systems. In: Proceed-
ings of the 13th International Conference on Broadband and Wireless Computing,
Communication and Applications (BWCCA-2018), pp. 444–455 (2018)
16. Saito, T., Nakamura, S., Enokido, T., Takizawa, M.: A causally precedent rela-
tion among messages in topic-based publish/subscribe systems. In: Proceedings of
the 21st International Conference on Network-Based Information Systems (NBiS-
2018), pp. 543–553 (2018)
17. Setty, V., van Steen, M., Vintenberg, R., Voulgais, S.: PolderCast: fast, robust,
and scalable architecture for P2P topic-based Pub/Sub. In: Proceedings of
ACM/IFIP/USENIX 13th International Conference on Middleware (Middleware-
2012), pp. 271–291 (2012)
18. Tarkoma, S.: Publish/Subscribe System: Design and Principles, 1st edn. John
Wiley and Sons, Ltd. (2012)
19. Tarkoma, S., Rin, M., Visala, K.: The publish/subscribe internet routing paradigm
(PSIRP): designing the future internet architecture. In: Future Internet Assembly,
pp. 102–111 (2009)
Data Analytics for Electricity Load
and Price Forecasting in the Smart Grid
Syeda Aimal1 , Nadeem Javaid1(B) , Amjad Rehman2 , Nasir Ayub1 ,

Tanzeela Sultana1 , and Aroosa Tahir3
1
2
Al Yamamah University, Riyadh, Saudi Arabia
3
Sardar Bahadur Khan Women University, Quetta 87300, Pakistan
Abstract. The present strategies for the prediction of price and load
may be difficult to deal with huge amount of load and price data. To
resolve the problem, three modules are incorporated within the model.
Firstly, the fusion of Decision Tree (DT) and Random Forest (RF) are
used for feature selection and to remove the redundancy among fea-
ture. Secondly, Recursive Feature Elimination (RFE) is taken for feature
extraction purpose that extracts the principle components and also used
for dimensionality reduction. Finally, to forecast load and price, Support
Vector Machine (SVM) and Logistic Regression (LR) as a classifiers are
used through which we achieve good accuracy results in load and price
prediction.
1 Introduction
Demand-side Management (DSM) is the main factor to motivate the user to

consume less energy during the peak hours, and also to shift the on-peak hours
energy load to off-peak hours. The main benefit for shifting the energy from on-
peak to off-peak hours is to making the smart grid cost-effective or profitable.
Smart grid is an electric grid that is consists of number of energy and oper-
ational measures like smart appliances, meters, and renewable energy resources
etc. The main objective of smart grid is to reduce electricity peak load and
to achieve the balance between power supply and demand [1]. Customers are
capable of partake inside the operations of smart grid, the cost of energy can be
minimized by the preservation of electricity and shifting the load. In this context,
dynamic pricing is a key indicator. The price forecasting is presumed because of
the requirement of financial system and enterprise [2]. As for customers, they’re
keen to know whether or not the price of electricity exceeds the certain con-
sumer defined thresholds, which they used to make decision to either turn the
load off or on. In these circumstances, clients require the electricity price classi-
fication. Hence, a few specific thresholds based totally on point price forecasting
algorithms are used to classify the price of electricity. Function approximation
https://doi.org/10.1007/978-3-030-15035-8_56
Electricity Load and Price Forecasting in Smart Grid 583
strategies are the fundamentals of forecasting algorithms, in which the basic

manner of rate formation is imitated by a price model [3]. Moreover, rate clas-
sification calls for decrease accuracy. Thus, the classification will become a key
precedence inside the price forecasting. The energy rate is influenced by the
usage of different factors, like with power requirement, renewable power supply
and it varies hourly.
1.1 Probem Statement

In paper [4,5], the authors claimed to improve price forecasting. However, the
load is not considered. The authors also worked on reducing the peak load
through different factors. However, the accuracy of load and price [6] is not
considered. The authors also made hybrid model of non-linear regression and
SVM. However, the load is not considered in this model. The electricity prices
are changing frequently, there is a need of a model which can over come these
challenges.
1.2 Paper Organization

Section 2, the literature of different related works is reviewed which is based
on feature engineering and electricity price and load forecasting. There are two
different ways for electricity load and price forecasting i.e. time-series model
and machine-learning model. Section 3, contains model in which the primary
goal of the model is to calculate the efficiency and accuracy. For this purpose,
we first need to collect the real data, preprocessing is done on that raw data,
selected features are extracted out and then parameters are tuned carefully.
Section 4, contains simulation and results in which we simulate our proposed
model, we performed the simulation in python environment. The simulations
are performed on system with specification intel corei5, 8 GB RAM, and 1 TB
hard disk. Section 5, contains conclusion of this paper.
2 Related Work
In the operations of modern electricity power system there is the choice of load
shifting as well as we can reduce the electricity energy cost. Dynamic electricity
pricing [7] and [8] helps the user to switch the load.
There are two main ways for the forecasting of load and price i.e. time-
series and machine learning model. For electricity load forecasting [9] developed
a model for the prediction of load and temperature of day ahead electricity
market with the analysis of singular spectrum and neural network structure.
The users are well pleased by knowing that the prices of electricity transcend
the thresholds, on the basis of which they can make the decision to whether turn
off or turn on the load. In the current situation, it is necessary to implement
price data analytic for the classification of electrical price and well as load. There
are various methods to apply classification for load and price forecasting i.e. by
584 S. Aimal et al.
categorizing the price on the basis of applied threshold values. As compared to

load, it is difficult to achieve accuracy in price due to factors influencing it.
Moreover, to come up with these issues, the authors applied various tech-
niques, traditional classifiers like Support Vector Machine widely known as SVM,
Navies Bayes, Neural Network, Decision Tree, and Random Forest are the most
famous and well-known one [10]. A survey is done in paper [11], according to
that survey we came to know that SVM outperform better accuracy but is high
in complexity. Like SVM, Logistic Regression is also a classifier that is used to
predict a binary outcome.
Time-series analysis is used largely in forecasting problem of price and
load. In [12], authors proposed an Auto-regressive Integrated Moving Average
(ARIMA) based algorithm for the forecasting purpose in the electricity market
of Turkey. The existence of different kinds of outliers in ARIMA, raw data that
is coming from the market. However, it is difficult to built a model with sta-
ble forecasting accuracy. In [13], an Auto-Regression Moving Average Hilbertian
(AMAH) is modeled for estimating the average values for time series which can
be applied to electrical price forecasting purpose.
For the application of the classifiers feature engineering is the bedrock for this
purpose. Feature engineering consist of two operations i.e. feature selection and
extraction. In load and price forecasting, different kinds of methods and tech-
niques are used for feature engineering. In [14], time-series data is processed by
using feature extraction. The extraction algorithm is based on Symbolic Aggre-
gation Approximation (SAX).
In the paper [15], for feature selection the author uses multi-variable mutual
information technique. In [16], the authors measures the feature’s relevancy.
Moreover, in [17] C4.5 algorithm for the selection of features which outperforms
better then Iterative Di-chotomiser 3 (ID3) in building the model of decision tree
[18]. Based on previous studies, the algorithms for feature selection or extraction,
classifier for modeling and tuning the parameters. The traditional classifiers are:
Artificial Neural Network (ANN), and DT are most widely used classifiers [19,
20]. However, in the case of DT, it often faces the outfit problem due to which it
does not performs good in forecasting however performs well in training. On the
other hand, ANN convergence cannot be controlled easily and its generalization
capability is so limited.
3 System Model
To enquire the problem of electricity load and price forecast. Our aim is to
predict the electricity load and price by using the data from grid and to achieve
good accuracy. To get over this issue, LR and SVM are taken as classifiers that
is used to predict the load as well as price. LR is a type of regression that is use
to predict the occurrence probability of an event, which is done when the data
is fitted to a logistic function. It may use several predictor variables that can be
either categorical or numerical. Historical data is used as a training-set on which
testing is done (Fig. 1). In the proposed approach, we first took the real data-set
from the electricity market (NYISO) i.e. is for three years from january 2011
till 2013. The required data is divided into month-wise and the data with the
same months are utilize for training and testing purpose. For example, for the
training purpose first three weeks of january from 2011 till 2013 are selected and
for the testing purpose last week of january 2011 till 2013 is selected. It consists
of following steps:
• First of all, for the suitable structure format it is necessary to format the
data.
• Incomplete and redundant values and variables are removed.
586 S. Aimal et al.
• The data sampling must be done to reduce consumption time taking by run-
ning algorithm.
The prepared data is used for training, validation, and testing. In this model,
feature extraction, selection, and classification is done. A fusion of DT, RF and
RFE is applied to the data-set for selecting and extracting the principle compo-
nents. RFE is used to eliminate the further redundancy in the selected features.
So, we can say that modules in this model consists of three parts: feature engi-
neering i.e. (selection, extraction) and classification.
The primary goal of the model is to calculate the efficiency and accuracy
while predicting the electricity load and price. For this purpose, we first took
real data from grid, selected features are extracted out and then parameters are
tuned carefully.
The following metrics are the most essential metrics for accessing the per-
formance of the model:
• The model that is applied for the purpose of predicting load and price should
work efficiently and run rapidly.
• Accuracy of classification is influence directly by the performance of feature
engineering.
• Classification’s accuracy is highly dependent on the goal of model.
The main issue that we are facing in the prediction of price and load is
accuracy due to the factors influencing them. However, it makes the training
tough. To enhance prediction accuracy, we embedded feature engineering process
(selection, extraction) in our model.
3.1 Feature Selector
In the feature selection, we combine DT and RF to select the important features

in data. A threshold value is given (i.e. 0.8) which selects the features that shows
the important features.
Feature selection for the following top reasons:
• The accuracy of a model can be improved.

• The complexity of a model can be improved.
• It increases the efficiency of machine learning algorithm to train faster.
3.2 Feature Extractor
After feature selection, feature extraction is done which eliminate the redun-
dancy among features using RFE.
• Improves the quality, speed, and effectiveness of supervised learning.

• It is a reduction process.
• it is used for the transformation of attributes.
3.3 Classifier
Classification is the method of discovering to which the set of new observation

belongs. This can be achieved on the basis of a trained data-set containing
observations. For the purpose of classification; LR and SVM classifiers are used
and tune them accordingly.
After eliminating the irrelevant and duplication in the features. Classifier is
applied for the prediction of load and price.
3.3.1 Steps for Logistic Regression

Feature scaling
• from sklearn.preprocessing import StandardScaler

• a = StandardScaler()
• train =a.fit.transform(train)
• test = a.transform(test)
Fitting Logistic Regression to dataset
• from sklearn.linear.model import LogisticRegression

• classifier = LogisticRegression()
• classifier.fit(X.train, y.train)
Predicting the test set result
• y.pred = classifier.predict(X.test)
Making the confusion matrix
• from sklearn.metrics import confusion.matrix

• cm = confusion.matrix(y.test, y.pred)
In this section, we are going to explain the following:
• Simulation Setup.
• Simulation Results.
• Evaluation Performance.
Each of the points mentioned above are describe in detail below.

588 S. Aimal et al.
4.1 Simulation Setup

To simulate our proposed model, we performed the simulation in python envi-
ronment. The simulation is performed on system with specification intel corei5,
8 GB RAM, and 1TB hard disk. Three years hourly electricity load data is used
for the evaluation of proposed scheme. The simulation results are organized as
follows:
• Features are selected with RF and DT.

• Redundant features are dropped also with RFE technique.
• Principle Components are selected.
• Logistic Regression performance is compared with four performance evalua-
tors i.e. Mean Average Error (MAE), Mean Square Error (MSE), Root Mean
Square Error (RMSE), and Mean Absolute Percentage Error (MAPE).

See Table 1.
Table 1. Attributes of data
Selected features Rejected features

DA-AA DA-AC
ST-BB RF-GE
SE-GEN SE-CC
RT-EC SE-MLC
– RT-CC
– RT-MLC
Fig. 2. Feature importance Fig. 3. Feature importance

Fig. 4. Load forecast for one month
4.2.1 Feature Selection Performance of DT and RE

DT and RF techniques are applied to select the important features with respect
to target feature from electricity load and price data i.e. for three years and the
RFE is applied to reduce the dimensions of data. The features, which have little
impact on load and price are dropped from data (shown in Figs. 2 and 3).“ In
Figs. 4 and 5, we see that LR as a classifier outperforms the best at the forecast
accuracy of electric load as well as price as compared to SVM. The accuracy
that is achieved by LR is 90%. However, if we see SVM it is not performing the
best as compared to LR. In load and price prediction of one month, we see that
LR is more close to the actual load and price as compared to SVM which is good
result. The normalized load of three years as shown in Fig. 7.
Fig. 5. Price forecast for one month

590 S. Aimal et al.
4.3 Performance Evaluation

For performance evaluation purpose, four means are used i.e. MAE, MSE,
RMSE, and MAPE. MAPE has the lowest error value i.e. 9.834 and MSE has
the highest error value i.e. 189.6475. In order to investigate the robustness and
accuracy of the techniques, the comparison is conducted as shown in Fig. 6.
Fig. 6. Error value comparison Fig. 7. Normalized load
5 Conclusions
In this paper, we’ve enquired the problem of prediction for electricity load and
price in smart grid. To solve the problem, forecasting model is used that is easy to
implement on distributed and parallelized systems. The proposed model is based
on feature engineering and classification. A fusion of two techniques is applied
for feature selection. To further remove the redundancy among features RFE
is used to extract the principle features. Now we have the new features which
enhances the classifiers and achieves the accuracy. LR as a classifier outperforms
best at the forecast accuracy of electric load as well as price. The accuracy that
is achieved by LR is 90%. However, if we see SVM it is not performing the best
as compared to LR. The performance is evaluated by comparing the results with
four means. i.e. MAPE, RMSE, MAE and MSE. For now, we have predicted
load and price for one month; in future we will attempt long term prediction.
i.e. for nine months as well as 12 months.
Acknowledgments. This research is supported by Al Yamamah university Riyadh

Saudi Arabia.
References
1. Zhang, D., Li, S., Sun, M., O’Neill, Z.: An optimal and learning-based demand
response and home energy management system. IEEE Trans. Smart Grid 7(4),
1790–1801 (2016)
2. Javaid, N., Hafeez, G., Iqbal, S., Alrajeh, N., Alabed, M.S., Guizani, M.: Energy
efficient integration of renewable energy sources in the smart grid for demand side
management. IEEE Access 6, 77077 (2018)
3. Jindal, A., Singh, M., Kumar, N.: Consumption-aware data analytical demand
response scheme for peak load reduction in smart grid. IEEE Trans. Ind. Electron.
(2018)
4. Wang, K., Xu, C., Guo, S.: Big data analytics for price forecasting in smart grids.
In: 2016 IEEE Global Communications Conference (GLOBECOM), pp. 1–6. IEEE
(2016)
5. Wang, K., Xu, C., Zhang, Y., Guo, S., Zomaya, A.: Robust big data analytics for
electricity price forecasting in the smart grid. IEEE Trans. Big Data (2017)
6. Moon, J., Kim, K.H., Kim, Y., Hwang, E.: A short-term electric load forecasting
on Big Data and Smart Computing (BigComp), pp. 219–226. IEEE (2018)
7. Keles, D., Scelle, J., Paraschiv, F., Fichtner, W.: Extended forecast methods for
day-ahead electricity spot prices applying artificial neural networks. Appl. Energy
162, 218–230 (2016)
8. Chen, P.C., Kezunovic, M.: Fuzzy logic approach to predictive risk analysis in
distribution outage management. IEEE Trans. Smart Grid 7(6), 2827–2836 (2016)
9. Mujeeb, S., Javaid, N., Akbar, M., Khalid, R., Nazeer, O., Khan, M.: Big data
analytics for price and load forecasting in smart grids. In: International Conference
on Broadband and Wireless Computing, Communication and Applications, pp. 77–
87. Springer, Cham (2018)
10. Javaid, N., Javaid, S., Abdul, W., Ahmed, I., Almogren, A., Alamri, A., Niaz,
11. Mahmood, D., Javaid, N., Ahmed, I., Alrajeh, N., Niaz, I.A., Khan, Z.A.: Multi-
agent-based sharing power economy for a smart community. Int. J. Energy Res.
41, 2074 (2017)
12. Zhao, Z., Lee, W.C., Shin, Y., Song, K.B.: An optimal power scheduling method for
demand response in home energy management system. IEEE Trans. Smart Grid
4(3), 1391–1400 (2013)
14. Ahmad, A., Javaid, N., Alrajeh, N., Khan, Z.A., Qasim, U., Khan, A.: A modified
feature selection and artificial neural network-based day-ahead load forecasting
model for a smart grid. Appl. Sci. 5(4), 1756–1772 (2015)
15. Ahmed, M.S., Mohamed, A., Khatib, T., Shareef, H., Homod, R.Z., Ali, J.A.: Real
time optimal schedule controller for home energy management system using new
binary backtracking search algorithm. Energy Build. 138, 215–227 (2017)
16. Ahmad, A., Javaid, N., Guizani, M., Alrajeh, N., Khan, Z.A.: An accurate and
fast converging short-term load forecasting model for industrial applications in a
smart grid. IEEE Trans. Ind. Inform. 13(5), 2587–2596 (2017)
17. Wu, M., Wang, Y.: A feature selection algorithm of music genre classification based
on ReliefF and SFS. In: IEEE International Conference on Computer and Informa-
tion Science (ICIS), pp. 539–544 (2009). Processing and Communications Appli-
cations, 2009, pp. 61–64
18. Wu, M., Wang, Y.: A feature selection algorithm of music genre classification based
on ReliefF and SFS. In: IEEE International Conference on Computer and Infor-
mation Science (ICIS), pp. 539–544 (2015)
19. Fleury, A., Vacher, M., Noury, N.: SVM-based multimodal classification of activ-
ities of daily living in health smart homes: sensors, algorithms, and first experi-
mental results. IEEE Trans. Inf. Technol. Biomed. 14(2), 274–283 (2010)
20. Huang, D., Zareipour, H., Rosehart, W.D., Amjady, N.: Data mining for electricity
price classification and the application to demand-side management. IEEE Trans.
Smart Grid 3(2), 808–817 (2012)
An Efficient CNN and KNN Data
Analytics for Electricity Load Forecasting
in the Smart Grid
Syeda Aimal1 , Nadeem Javaid1(B) , Tahir Islam1 , Wazir Zada Khan2 ,

Mohammed Y. Aalsalem2 , and Hassan Sajjad1
1
2
Farasan Networking Research Laboratory,
Department of Computer Science and Information System, Jazan University,
Jazan 82822-6694, Saudi Arabia
Abstract. The significant part of smart grid is to make smart grid

cost-efficient by predicting electricity price and load. To improve the
prediction performance we proposed an Efficient Convolutional Neural
Network (ECNN) and Efficient K-nearest Neighbour (EKNN) in which
the parameters are tuned. It may be difficult to deal with huge amount
of load data that is coming from the electricity market. To overcome
this issue, we incorporated three modules in the proposed methodology.
The proposed model consists of feature engineering and classification.
Feature engineering is a two-step process (feature selection and feature
extraction); for the purpose of feature selection Mutual Information (MI)
is used which reduces the redundancy among features and for feature
extraction Recursive Feature Elimination (RFE) is used to extract the
principle features from the selected features and reduces the dimension-
ality of features. Finally, after training the data-set and the removal of
the duplicate features load prediction is done by ECNN and EKNN. The
ECNN and EKNN outperforms better then traditional Convolutional
Neural Network (CNN) and K-nearest Neighbour (KNN). The forecast
performance is evaluated by comparing the results with MAPE, RMSE,
MAE and MSE. i.e. 10.8, 7.5, 7.15, and 10.4 respectively.
1 Introduction
Nowadays data analytic plays significant role in every field. Smart grids intro-
duced an advance informational infrastructure that allows optimization in energy
production, transmission, distribution and storage. Smart grid collects data from
different resources and stores it using data analytic techniques. In Smart grid,
bi-directional communication generates data having huge volume, variety and
velocity. Advance data analytic techniques are needed to figure out the required
information from the available raw data. Machine learning techniques are one

https://doi.org/10.1007/978-3-030-15035-8_57
Electricity Load Forecasting 593
of the reliable methods for predicting load and price. Supervised learning algo-
rithms are the part of machine learning which are used for predictions. Cate-
gories of machine learning algorithms are classification and regression. KNN and
CNN are the part of supervised learning algorithm. Deep learning is a learning
method for multiple levels of abstraction and representation which contains an
input layer, hidden layer, and an output layer. Deep learning comes under an
umbrella of machine learning which is based on learning levels of representation,
corresponding to a hierarchy of features, factors or concepts, where the concept
of higher-level is derived from lower-level and vice versa.
The recent advancements in sensor based technologies hold the promise of
novelty in energy consumption models that can better characterize the underly-
ing patterns. Demand-side Management (DSM) is the main and essential com-
ponent towards energy efficiency and management. DSM in smart homes and
residential buildings is a key component towards sustainability and efficiency in
urban environment.
Smart grid is an electrical grid which consists of range of energy and opera-
tional measures. The main objective of grid is to scale back electricity peak load
and to attain the balance between power supply and demand. Customers are
encouraged to partake within the operations of smart grid, the cost of electric
load will be reduced by the preservation of electricity and by shifting the load
towards off-peak hours.
1.1 Motivation or Problem Statement

In paper [1], the authors forecasted electricity price with deep learning
approaches and empirical comparison of traditional algorithms. Similarly, in
[3], the authors tried to forecast price. However, the load is neglected. Authors
also worked on peak load reduction. However, the accuracy of load and price is
neglected.
1.2 Contribution
In this paper, We have proposed new classification techniques (i.e. ECNN and
EKNN). The terms forecasting and prediction are used alternatively.
Load is predicted using the new classifiers which beats the traditional tech-
niques in terms of accuracy. Prediction is done on NYISO Data-set. We incorpo-
rated three modules in the proposed methodology. The proposed model consists
of feature engineering and classification. Feature engineering is a two-step pro-
cess (feature selection and feature extraction); for the purpose of feature selec-
tion Mutual Information (MI) is used which reduces the redundancy among fea-
tures and for feature extraction Recursive Feature Elimination (RFE) is used to
extract the principle features from the selected features and reduces the dimen-
sionality of features. Finally, after training the data-set and the removal of the
duplicate features load prediction is done by ECNN and EKNN. The ECNN and
EKNN outperforms better then traditional Convolutional Neural Network and
K-nearest Neighbour.
594 S. Aimal et al.
The main contributions are summarized as follow:
• The accuracy must be improved.

• Reduce the computational time.
• Less resources are utilized.
• Enhance the classifier to increase the prediction accuracy.
The rest of this paper is organized in the following way:
In Sect. 2, the related work in context of forecasting load and price is
reviewed. In Sect. 3, the system model is explained. In Sect. 4, the simulations
are performed in python environment on intel core i5, 8 GB RAM, and 1 TB
hard disk. In Sect. 5, the result of this paper is concluded.
2 Related Work
In this section, we have discussed different proposed schemes and techniques for
price and load forecasting [1]. The accuracy in forecasting helps the energy sup-
pliers to optimize their strategies for gaining profit. Likewise, to gain a successful
trade in electricity market different techniques have been proposed by different
authors. Short-term Price Forecasting (STPF) and Short-term Load Forecasting
(STLF) play principle role in the management of energy and resource trading
in the grids [2]. Similarly, Auto-regressive Integrated Moving Average (ARIMA)
performs good in stable electricity market.
In the electricity market, load and price forecasting are essential for optimal
operation planning; most of the load and price forecasting methodologies suffer
from lack of efficient feature selection techniques. To overcome this issue, the
authors in [3] proposed a new hybrid filter-wrapper approach that selects the
most relevant features according to the costumer input in a coordinated manner.
In [4], to construct a forecasting model authors presented a hybrid algorithm
which concatenate Empirical Mode Decomposition (EMD), Similar Days (SD)
Selection, and Long Short-term Memory (LSTM) neural network. Moreover,
to evaluate the accuracy of model between the forecasting and historical data
extreme gradient boosting-based weighted k-means algorithm is adopted.
The authors also deal with two types of forecasting online and day-ahead
hourly forecasting [5]. Stacked Denoising Auto-encoder (SDA) models are com-
pared with data-driven approaches in online forecasting whereas in day-ahead
forecasting, the accuracy and effectiveness is further validated. Mathematical
and statistical models like ARIMA, SARIMA etc., belongs to the classical type
of classifiers whereas artificially intelligent models are CNN, LSTM [6] (Table 1).
The authors in [7], proposed an Improved Elman Neural Network (IENN)
based forecast engine to forecast the load. The weights for this prediction model
carried out with an intelligent algorithm to find better forecasting results.
The efficiency of this model is brought-out to the real world engineering in
the comparison with other forecasting models [8]. In [9], the authors focused
on coordinated management of renewable and traditional energy in which the
Table 1. Summary of related work
Technique Objective Data Set Limitations

RBM, Relu Load PJM Overfitting problem
SNN, DNN Price AEMO Cannot deal with
non-linearity
RNN, LSTM Load IRISH Time complexity
CNN, LSTM Price International Overfitting problem,
Exchange Computationally expensive
C4.5, ID3 Price prediction, – Redundant feature not
C4.5 beats ID3 in removed
decision tree
building
Hilbertian Price EPEX Cannot deal with extreme
ARMAX non linearity
MLR, NN Load ISONE Not suitable for long term
forecasting
MI, ANN Load PJM Time complexity
LSTM, DNN, Load and price EPEX Over fitting
GRU
DRN Load and ISONE Over fitting
temperature
ANN, SVM, Load and price NSW Hard to tune parameters
RBFNN
DWT-IR, Load and price NYISO Slow convergence
SVM, Sperm
Whale
STLF, STPF Load and weather PJM, NYISO Computationally expensive
QO-ABC, Load and price NYISO, PJM Time complexity
LS-SVM
DNN Load and price EPEX Space complexity
DNN, LSTM Price and load NYISO Redundancy features not
prediction removed
authors formulate the problem as a three-stage Stackel-berg game in which the

authors employ backward induction and nash equilibrium method.
The authors also proposed unsupervised data clustering and frequent pattern
mining analysis on energy time series [10], and Bayesian network [10] prediction
for energy usage. The authors also perform experiments using real-world smart
meter data-sets. Accuracy in prediction is a considerable issue with multiple
data [11] which are applicable for a specified time frame. For each day ahead
coming-up with, the operations and the facility from the wind resources can be
foreseen. Observations and a series of models that span the temporal and spacial
596 S. Aimal et al.
scales of the matter; correct prediction could be a massive knowledge downside

that needs disparate knowledge. In [12], multiple models that are applicable for
a particular time-frame, and application of artificial intelligence techniques [13]
are studied.
In paper [14], the authors propose a new prediction methodology that makes
an attempt to generalize the quality of ARMAX statistical model to pair of
metric space. The structure of this planned model is a statistical regression where
parameters operate variables. The variables are lagged values of the series. i.e.
past determined innovations (like moving average terms), or exogenous variables.
In paper [15], a prediction strategy is planned for electricity markets. The
strategy uses large amount of data with hourly information; which is further
used as an input for two separate prediction models. Moreover, a rolling horizon
framework is planned to supply correct updates for the predictions.
The model in paper [16] is intended to find the incidence of abnormal events
and conditions. A generative model for anomaly detection is taken into consider-
ation and the hierarchical data-set of the network is collected from smart meters.
We tend to additionally address three challenges existing in smart grid:
1. large-scale variable count measurements

2. missing points
3. variable choice
In [17], the value prediction is entirely viewed as a symbol process. Data

integrity in [18], the attacks are expected to damage the performance of predic-
tion systems, which can have a significant impact on power systems and therefore
the resilience of power grids.
It reveals the result of knowledge integrity attacks on the accuracy of four repre-
sentative load forecasting models (multiple regression towards the mean, support
vector regression, artificial neural networks, and fuzzy interaction regression).
In paper [19], the authors tend to propose a probabilistic data-driven model
for consumption prediction in residential buildings. The model relies on theorem
network framework that is ready to find dependency relations between conduc-
tive variables which are able to relax the assumptions that are usually created
in ancient prediction models.
3 System Model
In the proposed approach, we enquire the problem of electricity load prediction.
In our prediction model, we tunned the hyper-parameters of KNN. In addition
to this, we also enhanced CNN in which we used two MAX-pooling and three
convolutional layers. Dense and flatten are used as a fully connected layers.
Historical load data that is taken from grid (NYISO) is used as a training data-
set for EKNN and ECNN; we used 80% of data for training purpose and rest of
20% is used for testing (Fig. 1).
Fig. 1. Prediction model
The proposed approach depicts that model is made up of three parts:
1. Feature selection
2. Feature extraction
3. Classification
Feature engineering which is two-step process (selection and extraction), and

then classification. In our prediction model, we first took the real data-set of
one year (i.e. january 2016–2017) from NYISO. The data is hourly-base which
is further utilize for the purpose of training and testing.
Design Goal: The goal of this model is to predict the load accurately and
efficiently. For this purpose, the preprocessed data is used on which feature
engineering process and classification is done.
To overcome these constraints:
• The data sampling must be done for the reduction of consumption time.
• Data formatting must be done carefully.
• Redundant data must be removed.
3.1 The Following Metrics Are the Most Essential Metrics

for Accessing the Performance of the Model
• Proposed model that is applied for the purpose of predicting load must work
efficiently.
• Accuracy of classification can be influenced directly by the performance of
feature engineering.
• Forecasted accuracy is highly dependent on the goal of model.
598 S. Aimal et al.
Algorithm 1. Prediction Algorithm

Input : Load Data
/* Separating target from features */
X : features of data
Y : target data
x train, x test, y train, y test = split(x,y);
combine imp = MI-imp + RFE-imp;
MI imp = importance calculate by MI;
selected feature = RFE(5, x train, y train);
IF
Combine imp of MI threshold And RFE is true, then
Reserve feature;
ELSE
Drop feature;
ENDIF
Prediction using EKNN and ECNN with tuned parameters;
Compare predictions with y test
3.1.1 Feature Selector

In the feature selection, the importance of features can be generated. In our
model we first set a certain threshold value i.e. 0.5 for the selection of important
features. In our model, MI is used for feature selection.
Feature selection is used for the following top reasons:
• It allows us to reduce over-fitting.

• The accuracy of a model can be improved.
• The complexity of a model can be improved.
• It increases the efficiency of machine learning algorithm to train faster.
3.1.2 Feature Extractor

When it comes to feature extraction, it is different from feature selection. After
the process of feature selection, feature extraction is done which down-sample
the most principle features. In our case, we extracted five principle features by
using RFE.
• It is used to further improve the quality, speed, and effectiveness of supervised

learning.
• It is a reduction process.
• it is used for the transformation of attributes.
3.1.3 Classifier
In Classification, we discover a new set of observations on the basis of a trained
set of data. For load prediction, we used ECNN and EKNN as a classifiers.
Algorithm 2. Algorithm of KNN

Input
Training data is loaded
Choose from the feature F
for Each point in data: do
Find Euclidean distance
Euclidean distance is stored in sorted list
The first k point is choosen
Assign a class on the basis of majority of classes
end forR(f-i + 1) - f*
End
Algorithm 3. Hyperparameter Optimization Algorithm

Input
Start procedure
Choose from the feature-set F
for TrainNetwork do
ModelEstimation
end forReturnBestHyperparameters( )
End

In this section, we are going to explain the following:
• Setup.
• Results.
• Evaluation Performance.
Each of the points mentioned above are describe in detail below.

To simulate our proposed model, we performed the simulation in python envi-
ronment. The simulation is performed on system with specification intel core i5,
8 GB RAM, and 1 TB hard disk. One year hourly electricity load data of NYISO
is used for the evaluation of proposed scheme.

• Features are selected with MI.
• Principle Components are extracted with RFE.
• Actual load is compared with ECNN and EKNN.
• The performance is compared with four performance evaluators i.e. Mean
Average Error (MAE), Mean Square Error (MSE), Root Mean Square Error
(RMSE), and Mean Absolute Percentage Error (MAPE).
600 S. Aimal et al.
Fig. 2. Importance feature selection with MI
Fig. 3. Feature extraction with RFE

Fig. 4. Load predicted by ECNN and EKNN
MI Technique is applied to select the importance of features with respect to

the target on the basis of a certain threshold value (i.e. 0.5) (Fig. 2). Features
which have little impact on load are dropped from data as shown in Fig. 4. The
most five principle feature are extracted to reduce the features dimensionally
through RFE as shown in Fig. 3. ECNN and EKNN as a classifier outperforms
best at the forecasting accuracy of electric load as shown in Fig. 4. The accuracy
that is achieved by ECNN is 92% and the accuracy that is achieved by EKNN
is 93%. The ECNN and EKNN is close to the actual load which is good result.
In order to investigate the robustness and accuracy of the proposed model; a

comparison is conducted as shown in Fig. 5. For performance evaluation, four
means are used. i.e. MAE, MSE, RMSE, and MAPE. MAPE whose error values
are i.e. 10.847888888888887, 7.566138821530217, 7.155807083333333, and 10.424
respectively.
TM
1 Av − Fv
M AP E = | | ∗ 100 (1)
T tm=1 Av

TM
1
RM SE = (Av − Fv )2 (2)
T tm=1
TM
1
M SE = (Av − Fv )2 (3)
T tm=1
N
n=1 |(Fv − Av )|
M AE = (4)
N
602 S. Aimal et al.
Fig. 5. ECNN performance evaluation metrics
5 Conclusions
In this paper, the problem of electricity load forecast is taken into consideration.
The data is obtained from grid (NYISO data-set) for one year (i.e. january 2016
to 2017). We took required data from the grid which is hourly-base data that
is further utilize for the purpose of training and testing. The process of feature
engineering is done in which the important features are selected using MI. After
the selection of important features, RFE is used for extracting five principle
features.
The trained data-set is further used for load prediction that is done by our
proposed classifiers (ECNN accuracy is 92% and for EKNN accuracy is 93%).
The predicted electricity load outperforms better then other classifiers in terms
of accuracy. The forecast performance is evaluated by comparing the results with
MAPE, RMSE, MAE and MSE. i.e. 10.8,7.5, 7.15, and 10.4 respectively.
References
1. Lago, J., De Ridder, F., De Schutter, B.: Forecasting spot electricity prices: deep
learning approaches and empirical comparison of traditional algorithms. Appl.
Energy 221, 386–405 (2018)
2. Liu, B., Nowotarski, J., Hong, T., Weron, R.: Probabilistic load forecasting via
quantile regression averaging on sister forecasts. IEEE Trans. Smart Grid 8(2),
730–737 (2017)
3. Wang, K., Xu, C., Zhang, Y., Guo, S., Zomaya, A.: Robust big data analytics for
electricity price forecasting in the smart grid. IEEE Trans. Big Data (2017)
4. Zheng, H., Yuan, J., Chen, L.: Short-term load forecasting using EMD-LSTM neu-
ral networks with a Xgboost algorithm for feature importance evaluation. Energies
10(8), 1168 (2017)
5. Ahmad, A., Javaid, N., Alrajeh, N., Khan, Z.A., Qasim, U., Khan, A.: A modified
feature selection and artificial neural network-based day-ahead load forecasting
model for a smart grid. Appl. Sci. 5(4), 1756–1772 (2015)
7. Vrablecová, P., Ezzeddine, A.B., Rozinajová, V., Šárik, S., Sangaiah, A.K.: Smart
grid load forecasting using online support vector regression. Comput. Electr. Eng.
65, 102–117 (2018)
8. Liu, Y., Wang, W., Ghadimi, N.: Electricity load forecasting by an improved fore-
cast engine for building level consumers. Energy 139, 18–30 (2017)
9. Javaid, N., Hafeez, G., Iqbal, S., Alrajeh, N., Alabed, M.S., Guizani, M.: Energy
efficient integration of renewable energy sources in the smart grid for demand side
management. IEEE Access 6, 77077 (2018)
10. Singh, S., Yassine, A.: Big data mining of energy time series for behavioral analytics
and energy consumption forecasting. Energies 11(2), 452 (2018)
11. Abedinia, O., Amjady, N., Zareipour, H.: A new feature selection technique for load
and price forecast of electrical power systems. IEEE Trans. Power Syst. 32(1), 62–
74 (2017)
12. Javaid, N., Javaid, S., Abdul, W., Ahmed, I., Almogren, A., Alamri, A., Niaz,
13. Yu, C.N., Mirowski, P., Ho, T.K.: A sparse coding approach to household electricity
demand forecasting in smart grids. IEEE Trans. Smart Grid 8(2), 738–748 (2017)
14. González, J.P., San Roque, A.M., Pérez, E.A.: Forecasting functional time series
with a new Hilbertian ARMAX model: application to electricity price forecasting.
IEEE Trans. Power Syst. 33(1), 545–556 (2018)
15. Chitsaz, H., Zamani-Dehkordi, P., Zareipour, H., Parikh, P.P.: Electricity price
forecasting for operational scheduling of behind-the-meter storage systems. IEEE
Trans. Smart Grid 9(6), 6612–6622 (2018)
16. Moghaddass, R., Wang, J.: A hierarchical framework for smart grid anomaly detec-
tion using large-scale smart meter data. IEEE Trans. Smart Grid 9(6), 5820–5830
(2018)
17. Singh, N., Mohanty, S.R., Shukla, R.D.: Short term electricity price forecast based
on environmentally adapted generalized neuron. Energy 125, 127–139 (2017)
18. Luo, J., Hong, T., Fang, S.C.: Benchmarking robustness of load forecasting models
under data integrity attacks. Int. J. Forecast. 34(1), 89–104 (2018)
19. Bassamzadeh, N., Ghanem, R.: Multiscale stochastic prediction of electricity
demand in smart grids using Bayesian networks. Appl. Energy 193, 369–380 (2017)
A New System for Management of IoT
Sensors Considering Agile-Kanban
Keita Matsuo1(B) , Takeru Kurita2 , and Leonard Barolli1

1
{kt-matsuo,barolli}@fit.ac.jp
2
Student of Information and Communication Engineering,
s15b2019@bene.fit.ac.jp
Abstract. Recently, IoT devices and sensors are rapidly increasing in

the word. They are connecting to the Internet and are generating vast
packets in the network. Thus, the networks could be congested and the
performance will degrade. For this reason, it is important to decrease
the number of transmitted packets. Agile is a technique to develop the
software and manage the work efficiency. Kanban is a method to support
Agile development. In this work, we consider the use of Agile and Kanban
for management and control of IoT sensors.
1 Introduction
Recently, many communication systems have been developed in order to support
humans. Especially, IoT technologies, IoT devices and sensors can provide many
services for humans. The IoT sensors can be embedded in electronic devices. For
instance, if a refrigerator has some IoT sensors inside, it can tell us the ingredi-
ents for cooking through the Internet anytime and anywhere. The temperature
of the air conditioner in the room can be controlled. Also, television, cooking
heater, microwave, bath-unit, toilet-unit, light and so on are going to provide
new services.
Currently, the vehicles can be connected to the Internet. Toyota uses the word
“Connected Car System”. This system can collect the data using mobile network
and sensors in the car. These are different kinds of data such as the data taken
by the car when is moving, car speed, braking or acceleration, weather condition
and traffic jam information. These data are gathered to the datacenter through
the network by using Internet technologies. The data can be analyzed and used
for solving some traffic problems. There are also other applications using IoT
such as the use of IoT technologies for agriculture, factories, offices, schools,
hospitals and so on.
Because the IoT devices and sensors generate a huge amount of packets, the
network load is increased and the sensor lifetime decreases. The lifetime of O2
https://doi.org/10.1007/978-3-030-15035-8_58
A New System for Management of IoT Sensors Considering Agile-Kanban 605
sensors is in the range of 4 to 8 years and ozone sensors is in the range of 1

to 2 years. In order to increase the lifetime of these kind of sensors some new
systems and approached are needed.
We consider the use of Agile and Kanban (the meaning of Kanban in Japanese
is sign board) for management of IoT sensors. Agile is software developing
method that denotes “the quality of being agile, readiness for motion, nimble-
ness, activity, dexterity in motion”. The software development methods can offer
the answer to the business community asking for lighter weight along with faster
and nimbler software development processes [1,5]. Kanban is one of methods that
can support Agile process [11]. In this work, we propose a sensor management
system for increasing the lifetime of IoT sensors.
The structure of this paper is as follows. In Sect. 2, we introduce the IoT and
Kanban. In Sect. 3, we present our proposed IoT Sensors Management System
with Agile-Kanban. In Sect. 4, we show the conclusions and future work.
2 IoT and Kanban
2.1 IoT
Our life style is drastically changing by IoT technologies. The IoT can connect
various things to the Internet such as household electronics appliances, vehicles,
robots, communication devices, and some application softwares. Thus improving
our quality of life. In Fig. 1, we show the image of IoT environment. IoT sys-
tem has many sensors, which can be used for operating factories, farms, offices,
houses, and so on. However given the fact that most of IoT sensors are resource
limited and operate on batteries, the consumption power and life time of sensors
are important issues for wireless IoT sensor networks design [9,12].
Fig. 1. Image of IoT environment.

606 K. Matsuo et al.
There are some approaches for decreasing the number of packets in the net-
works. The Opportunistic Networks (OppNets) is one of them. It can provide
an alternative way to support the diffusion of information in special locations
in a city, particularly in crowded spaces where current wireless technologies can
exhibit congestion issues [8].
2.2 Kanban
Kanban is a system produced by Toyota. The system can efficiently produces

vehicles. The main goal of Kanban was to produce cars for the same or lower
price than the competitors. This Kanban system is called Toyota Production
System (TPS). The TPS is also known as Just in Time (JIT) manufacturing
and the basic principle is to produce “only what is needed, when it is needed
and in the amount needed”. Currently, there are many research work using Kan-
ban for developing software [6,7,13]. At the beginning, the Kanban was used in
manufacturing processes, however its applications in other areas is continuously
growing due to its proven successfulness.
We have shown a schematic illustration of Kanban system in Fig. 2. The
system has two Kanbans. The Kanban system has 3 sections: Upstream, Down-
stream and Store. The Upstream section role is for manufacturing parts by mak-
ing some blocks of products. Downstream section is used for assembling parts
and completing the products. Store section is used for keeping some parts needed
for working operation. If the amount of stocking parts increases, the manufac-
turing efficiency decreases. In order to solve this problem, Kanban system uses
Production-Kanban and Withdraw-Kanban.
The upstream section produces some parts and stock them in the store
section. While, the downstrean section uses the parts in the store to assem-
ble the products. The Production-Kanban can only move between upstream and
Fig. 2. Structure of Kanban system.

store, while the Withdraw-Kanban is moving between store and downstream as

shown in Fig. 2.
We describe the moving of Kanbans in Fig. 3. If there is a shortage of parts
in downstream section, the Withdraw-Kanban moves from downstrean section
to store section. Then, the Withdraw-Kanban informs the number of shortage
parts to Production-Kanban. After that, the Production-Kanban instructs the
upstream section to manufacture the number of shortage parts informed by
Withdraw-Kanban. The Production-Kanban will put the number of shortage
parts in the production queue. Then, the production of parts will starts. Thus,
the Kanban system can control the supply and demand of parts, which leads to
an efficient manufacturing.
Fig. 3. Moving of Production-Kanban and Withdraw-Kanban.
Recently, Kanban system combined with Agile is used for software devel-
opment. Some papers use Kanban and Agile for collaboration work [3,10]. In
another paper, the authors use Agile approach with Kanban for managing the
security risk on e-commerce [2].
3 Proposed System
We propose IoT sensor management system considering Agile-Kanban (see

Fig. 4). Our proposed system uses Kanboard. The Kanboard requires the web
server, database and PHP. The Kanboard offers 4 kinds of Kanbans to users
which are Backlog, Ready, Work in progress and Done. Each Kanban descrip-
tion is shown in Table 1.
Fig. 4. Proposal system with Agile-Kanban.
Table 1. Description of Kanban names.
Name of Kanban Description

Backlog The sensor needs maintenance
Ready The sensor is ready for using
Work in progress The sensor is working
Done The sensor finish the work
3.1 Overview of Kanboard
Kanboard is a free and open source software [4]. Kanboard user interface is
shown in Fig. 5. The Kanboard focus on simplicity minimalism. The number of
features is limited. By Kanboard can be known the current status of a project
because it is visual. It is very easy to understand and there is no need to explain
and no training is required.
Fig. 5. User interface of Kanboard.

Kanboard has a number of features as follows.

• Action Visualization,
• Limit work in progress to focus on the goal,
• Drag and drop tasks to manage the project,
• Self-hosted,
• Simple installation.
In Fig. 5, we present a user interface of Kanboard, which shows 4 Kanbans:

Backlog, Ready, Work in progress and Done.
3.2 IoT Sensors Management System

Our proposed IoT sensors management system can manage various kinds of
sensors. The Kanbans can move in different states such as Backlog, Ready, Work
in progress and Done. One Kanban corresponds to one sensor. The Kanban can
change some states with drag and drop. For example, when there is a Kanban in
the state of Backlog, it means that the sensor needs to do maintenance. When
Kanban is in the state of Ready, it means that the sensor is ready to be used.
Work in progress means the sensor is working. In this case, the measurement
data by sensor will be uploaded to the database on the Internet. When Kanban
is in the state of Done, it means the end of work.
In Fig. 6 is shown the user interface of IoT management system. When the
Kanban is in the “Work in progress” state, the measurement data of temperature
and humidity are shown in Fig. 7. In Fig. 8 is shown the image of moving Kanban
on the IoT sensor management system using Kanboard.
The benefits of our proposed IoT sensor management system are as following.
• The system can manage a large amount of IoT sensors.
• When the IoT sensors are not used, the management system can cut down
the consumption power for saving IoT sensor’s battery.
• Only when the users are using the IoT sensor, the sensors are able to com-
municate through the network. This means that the system can reduce the
wast packets.
Fig. 6. User interface of IoT management system.

Fig. 7. The image of measured data on the Kanban.
Fig. 8. Different states of Kanban moving.
• The system is able to analyze the IoT sensors status such as sensor’s working
time, life time, frequency of use and sensor’s activity.
• The proposed system can be installed in factory machines, vehicles, robots,
offices or school facilities, home electrical appliances and so on.
In this paper, we introduced Agile-Kanban. We presented in details the Kanban

system and Kanboard. We proposed a new system for management of IoT sen-
sors considering Agile-Kanban. The system can manage a large amount of IoT
sensors. Our proposed system can decrease the IoT sensor’s consumption power,
thus increasing the IoT sensor lifetime. By making an efficient management, the
communication traffic is decreased.
In the future work, we would like to improve the proposed system and carry
out experiments.
References
1. Abrahamsson, P., Salo, O., Ronkainen, J., Warsta, J.: Agile software development
methods: Review and analysis. arXiv preprint arXiv:1709.08439 (2017)
2. Dorca, V., Munteanu, R., Popescu, S., Chioreanu, A., Peleskei, C.: Agile approach
with kanban in information security risk management. In: 2016 IEEE International
Conference on Automation, Quality and Testing, Robotics (AQTR), pp. 1–6. IEEE
(2016)
3. Hofmann, C., Lauber, S., Haefner, B., Lanza, G.: Development of an agile develop-
ment method based on kanban for distributed part-time teams and an introduction
framework. Procedia Manuf. 23, 45–50 (2018)
4. Kanboard: Kanboard web page. https://kanboard.org/
5. Kiely, G., Kiely, J., Nolan, C.: Scaling agile methods to process improvement
projects: a global virtual team case study (2017)
6. Kirovska, N., Koceski, S.: Usage of kanban methodology at software development
teams. J. Appl. Econ. Bus. 3(3), 25–34 (2015)
7. Maneva, M., Koceska, N., Koceski, S.: Introduction of kanban methodology and
its usage in software development (2016)
8. Miralda, C., Donald, E., Kevin, B., Keita, M., Leonard, B.: A delay-aware fuzzy-
based system for selection of IOT devices in opportunistic networks considering
number of past encounters. In: Proceedings of the 21th International Conference
on Network-Based Information Systems (NBiS-2018), pp. 16–29 (2018)
9. Nair, K., Kulkarni, J., Warde, M., Dave, Z., Rawalgaonkar, V., Gore, G., Joshi,
J.: Optimizing power consumption in IOT based wireless sensor networks using
bluetooth low energy. In: 2015 International Conference on Green Computing and
Internet of Things (ICGCIoT), pp. 589–593. IEEE (2015)
10. Padmanabhan, V.: Functional strategy implementation-experimental study on
agile kanban. Sumedha J. Manag. 7(2), 6–17 (2018)
11. Petersen, J.: Mean web application development with agile kanban (2016)
12. Sheng, Z., Mahapatra, C., Zhu, C., Leung, V.: Recent advances in industrial wire-
less sensor networks towards efficient management in IOT. IEEE Access 3, 622–637
(2015)
13. Yacoub, M.K., Mostafa, M.A.A., Farid, A.B.: A new approach for distributed soft-
ware engineering teams based on kanban method for reducing dependency. JSW
11(12), 1231–1241 (2016)
A Two-Way Flow Model for Fog
Computing
Yinzhe Guo1(B) , Ryuji Oma1 , Shigenari Nakamura1 , Dilawaer Duolikun1 ,

Tomoya Enokido2 , and Makoto Takizawa1
1
{yinzhe.guo.8m,ryuji.oma.6r}@stu.hosei.ac.jp,
nakamura.shigenari@gmail.com, dilewerdolkun@gmail.com,
2
eno@ris.ac.jp
Abstract. In our previous studies, the TBFC (Tree-Based Fog Comput-

ing) model of the IoT is proposed to reduce the electric energy consumed
by fog nodes and servers. Here, fog nodes are hierarchically structured
in a height-balanced tree. A root node shows a cloud of servers and leaf
nodes indicate edge fog nodes which communicate with sensors and actu-
ators. First, sensor data is sent to edge fog nodes. Each fog node receives
data from child fog nodes, processes the data, and sends the processed
data to a parent fog node. Finally, servers receive sensor data processes
by fog nodes and send actions to actuators. Thus, sensor data is sent
upwards to a root node from sensors and action are sent downward to
actuators from the root node. In order to more promptly deliver actions
to actuators, we newly propose a TWTBFC (Two-Way TBFC) model
where fog nodes not only send processed data to a parent fog node but
also send both the processed data and data received from the child fog
node to each child fog node. Thus, each fog node can get more global
information, i.e. not only data from child fog nodes but also data from
sibling fog nodes. Fog nodes nearer to edge fog nodes decide on actions
by using the data and send the actions to child fog nodes.
Keywords: Energy-efficient fog computing ·

IoT (Internet of Things) · Tree-based fog computing (TBFC) model ·
Two-Way TBFC (TWTBFC) model
1 Introduction
The IoT (Internet of Things) is composed of not only computers like servers
and clients but also devices like sensors and actuators installed in various things
[4]. Compared with traditional network systems like the cloud computing model
[1], the IoT is more scalable and huge amount of data from sensors are trans-
mitted in networks and are processed in servers. Networks and servers are too
heavily loaded to timely decide on actions for sensor data and deliver actions
https://doi.org/10.1007/978-3-030-15035-8_59
A Two-Way Flow Model for Fog Computing 613
to actuators. In order to efficiently realize the IoT, the fog computing model is
proposed [7].
The fog computing model [7] is used to reduce the network and server traffic
of the IoT (Internet of Things). Since the IoT is scalable, huge amount of electric
energy is consumed in the IoT [2,3]. In order to not only increase the performance
but also reduce the electric energy consumption of the IoT, the TBFC (Tree-
based Fog Computing) model is proposed in our previous studies [4–6]. Here,
fog nodes are hierarchically structured in a height-balanced tree. A root node
shows a cloud composed of servers and leaf nodes are edge nodes which receive
sensor data from sensors and send actions to actuators. Each fog node has one
parent fog node and at least one child fog node. Each sensor first sends sensor
data to an edge fog node. The edge fog node is a parent of the sensor and the
sensor is a child of the edge fog node. Each fog node receives input data from
child fog nodes. Then, the fog node processes the data and sends the processed
data i.e. output data to a parent fog node. Finally, servers in a cloud receive
data processed by fog nodes. Then, the servers make a decision on actions to
be done by actuators and deliver the actions to actuators through fog nodes.
Thus, sensor data is sent upwards from sensors to a cloud of servers. Servers
in the cloud send actions downward to actuators. Here, it takes time to deliver
actions to actuators because servers make a decision on actions to be done by
the actuator.
In this paper, we newly propose a TWTBFC (Two-Way TBFC) model so
that actions to be performed on actuators are decided by fog nodes nearer to
devices in order to more promptly deliver actions to actuators. Here, each fog
node not only sends output data obtained by processing input data to a parent
fog node in a same way as the TBFC model but also forwards every input data
as well as the output data to the child fog nodes. Thus, each fog node can get
more global information by knowing output data sent by other sibling fog nodes
and the decision of the parent fog node. For example, a fog node of each house
can get not only data like energy consumption in the house but also neighbors
in a town. By using the data, each fog node can decide on actions to be executed
by descendant actuators and can send the actions downward to actuators. We
also present an application of a home energy control system in the TWTBFC
model.
In Sect. 2, we present a system model. In Sect. 3, we propose the TWTBFC
model. In Sect. 4, we present an application of the home energy control system
of the TWTBFC model.
2 System Model
2.1 Fog Computing Model
The fog computing (FC) model [7] to realize the IoT [4] is composed of devices,
fog nodes, and clouds. Clouds are composed of servers like the cloud computing
(CC) model [1]. Various devices, i.e. sensors and actuators are connected to edge
614 Y. Guo et al.
fog nodes. A sensor sends sensor data to edge fog nodes. Fog nodes are inter-
connected with other fog nodes in networks. An application process is composed
of subprocesses. A subprocess of an application process to handle sensor data is
performed on a fog node to process a collection of input data sent by sensors
and other fog nodes. Then, new output data, i.e. processed data is generated by
a fog node through performing the subprocess on the input data. For example,
a minimum value vk is selected by searching a collection of input data v1 , ...,
vl obtained from sensor s1 . The minimum value vk is the output data and a
collection of sensor data v1 , ..., vl are the input data of the fog node. Output
data processed by a fog node is sent to another fog node. Thus, sensor data
are processed by fog nodes and forwarded to fog nodes. Servers finally receive
data processed by fog nodes. Actions decided by servers are sent to edge fog
nodes and the edge fog nodes issue the actions to actuator nodes. On receipt of
an action, an edge fog node sends the action to actuators. Thus, subprocesses
are performed on not only servers but also fog nodes in the FC model while
centralized to servers in the cloud computing model.
Fig. 1. TBFC model.
2.2 TBFC Model

In the tree-based fog computing (TBFC) model [5,6], fog nodes are structured in
a height-balanced tree as shown in Fig. 1. Here, the root node f0 denotes a cloud
of servers. Fog nodes at the bottom level are edge fog nodes which communicate
with sensor and actuators.
Each fog node fR has lR (≥ 0) child fog nodes fR0 , fR1 , ..., fR,lR −1 . Here,
fRi is the ith child fog node of the fog node fR and in turn fR is a parent fog
node of each fog node fRi . Let ch(fR ) show a set {fR0 , fR1 , ..., fR,lR −1 } of
child fog nodes of a fog node fR . Let pt(fRi ) denote a parent fog node fR of a
child fog node fRi . If a fog node fR is a root node of the tree, R = 0, i.e. fR
is f0 . If a fog node fR is an ith child fog node of a fog node fR , fR is fR i ,
i.e. R = R i. For example, the third child fog node of a fog node f010 is f0102 .
Thus, the label R of a fog node fR shows a path from the root node f0 to the
fog node fR . For the length |R| of the label R of a fog node fR , the level of the
fog node fR is |R| − 1. For example, the levels of the fog nodes f0 and f011 are
|0| − 1 = 1 − 1 = 0 and |011| − 1 = 3 − 1 = 2, respectively. If the label R is a
concatenation R1 R2 of labels R1 and R2 , a fog node fR1 is an ancestor of the
fog node fR . For example, a fog node f021 is an ancestor of a fog node f02101
and a parent fog node of f01213 .
Let as(fR ) be a set of ancestor fog nodes of a fog node fR , i.e. {fR1 | R =
R1 R2 for some labels R2 }. Let ds(fR ) be a set of descendant fog nodes of a fog
node fR , {fR1 | R1 = RR2 for some labels R2 }. A least upper bound lub(fR , fS )
of a pair of fog nodes fR and fS is a common ancestor fog node fT whose any
descendant fog node is not an ancestor of the fog nodes fR and fS . SG(fR ) is a
set of fog nodes which are at the same level as a fog node fR .
Fig. 2. Fog node of the TBFC model [6].
A fog node fR takes input data dRi from each child fog node fRi (i = 0, 1,
..., lR − 1). DR shows a collection of input data dR0 , dR1 , ..., dR,lR −1 of the fog
node fR . A subprocess p(fR ) in the fog node fR does the computation on the
input data DR and then generates output data dR . Then, the fog node fR sends
the output data dR to a parent fog node pt(fR ). A collection DR of the input
data dR0 , dR1 , ..., dR,lR −1 and output data dR are stored by the storage module
SR in the storage DBR .
A notation |d| shows the size [Byte] of data d. The ratio |dR |/|DR | is the
output ratio ρR of a fog node fR . Suppose DR is a set {v0 , v1 , v2 , v3 } of four
numbers showing temperature obtained by child fog nodes fR0 , fR1 , fR2 , and
fR3 , respectively. If the output data dR is a maximum value v of the values
v0 , v1 , v2 , and v3 , the output ratio ρR of the fog node fR is |dR |/|DR | = 1/4.
616 Y. Guo et al.
Here, ρR ≤ 1 and the size of output data dR is smaller than the input data DR .
Suppose input data dRi from each fog node fRi is a set of values. If the output
data dR is obtained by taking the direct product of the input data dR0 , dR1 , ...,
dR,lR −1 the size |dR | of the output data dR is |dR0 | · |dR1 | · ... · |dR,lR −1 |. Here,
ρR ≥ 1.
A fog node fR is composed of input (IR ), computation (CR ), storage (SR ),
and output (OR ) modules as shown in Fig. 2 [6].
3 A Two-Way TBFC (TWTBFC) Model

3.1 Two-Way Data Transmission
We propose a Two-Way TBFC (TWTBFC) model in this paper in order to more
promptly deliver actions to actuators. Fog nodes are hierarchically structured in
a height-balanced tree as discussed in the TBFC model. A root node f0 denotes
a cloud of servers. A fog node fR has lR (≥ 0) child fog nodes fR0 , fR1 , ...,
fR,lR −1 . An edge fog node has no child fog node and communicates with devices,
i.e, receive sensor data from sensors and sends actions to actuators. An edge fog
node fR has child sensors sR0 , sR1 , ..., sR,slR −1 (slR ≥ 0) and child actuators
aR0 , aR1 , ..., aR,alR −1 (alR ≥ 0). Every edge fog node is at the same level. Each
fog node fR of the same level is equipped with a same subprocess p(fR ).
Each sensor sRi sends sensors data udRi to a parent edge fog node fR . The
edge fog node fR first collects sensor data U DR = {udR0 , udR1 , ..., udR,slR −1 }
from child sensors sR0 , sR1 , ..., sR,slR −1 , respectively. Then, the edge fog node
fR sends output data udR obtained by processing the sensor data U DR to a
parent fog node pt(fR ) in a same way as the TBFC model.
A fog node fR receives input data udR0 , udR1 , ..., udR,lR −1 from child fog
nodes fR0 , fR1 , ..., fR,lR −1 . Then, the fog node fR generates output data udR
by performing the subprocess p(fR ) on the input data U DR . Then, the fog node
fR sends the output data udR to a parent fog node pt(fR ). Thus, data processed
by fog nodes is finally delivered to the wot node, i.e. sensor f0 as discussed in
the TBFC model.
In the TWTBFC model, each fog node fR forwards the upward output data
udR and the upward input data U DR to each child fog node fRi in addition to
sending the upward output data udR to a parent fog node. On receipt of the
upward input data U DR and upward output data udR from a parent fog node
fR , each child fog node fRi can know what upward output data udRj another
child fog node fRj (j = i) is generated. By using the upward output data udR0 ,
..., udR,lR −1 from the other sibling fog nodes, each child fog node fRi can make
some decision on actions to be done by descendant actuators. The fog node fRi
sends the downward output data DDRi including the upward input data udRi0 ,
..., udRi,lRi −1 and the decision, i.e. downward output data ddRi to each child fog
node fRij as shown in Fig. 3.
If fR is an edge fog node, the downward output data DDR is a set of
actions acR0 , acR1 , ..., acR,dR −1 on the child actuators aR0 , aR1 , ..., aR,alR−1 ,
respectively. Here, each action acRi is an action to be done on an actuator aRi
Fig. 3. TWTBFC model.
(i = 0, ..., alR − 1). Each edge fog node fR sends the actions acR0 , acR1 , ...,
acR,lR −1 to the child actuators aR0 , aR1 , ..., aR,alR −1 , respectively.
In the TWTBFC model, each fog node fR has two computation modules
CUR and CDR to process data U DR from child fog nodes fR0 , ..., fR,lR −1 and
to process data from the parent fog node pt(fR ).
3.2 Fog Nodes

We consider modules of each fog node fR in the TWTBFC model. As shown in
Fig. 4, an input module IUR of a fog node fR receives upward input data udRi
from each child fog node fRi (i = 0, ..., fR,lR −1 ). Then, an upward computation
module CUR does the computation on the upward input data U DR and generates
upward output data udR . An upward output module OUR sends the upward
output data udR to a parent fog node fR ( = pt(fR )). The upward IUR , CUR ,
and OUR modules are same as the input, computation and output modules of
the TBFC model.
In addition, the fog node fR receives downward input data DDR from a
parent fog node fR ( = pt(fR )). DDR is composed of upward input data U DR
and upward output data udR of the parent fog node fR . The downward input
module IDR receives the downward input data DDR from the parent fog node
fR . The downward input data DDR is processed by the downward computation
module CDR with the upward input data U DR and the upward output data udR
and generates downward output data DDR . The downward computer module
618 Y. Guo et al.
Fig. 4. Fog node of the TWTBFC model.
CDR makes some decision on actions to be executed by descendant actuators.

The downward output module ODR sends not only DDR but also U DR to each
child fog node fRi .
4 Example
We consider an electric energy management system at home as an example of
the TWTBFC model. There are devices which supply electric energy to home
and consume electric energy at home. We consider the following five types of
devices:
1. Home (d0 ).
2. Rechargeable battery (RB) (d1 ).
3. Solar panel (SP) (d2 ).
4. Fuel cell (FC) (d3 ).
5. Electric company (EC) (d4 ).
Devices are classified in terms of two points, supplier and consumer of electric
energy. The home device d0 is just a consumer type, i.e. consumes electric energy.
The rechargeable battery (RB) device d1 is characterized by both consumer and
supplier type. The RB (d1 ) device supplies electric energy to the home device
(d0 ) but also consumes electric energy from the EC, SP, and FC devices. The
solar panel (SP) device d2 and fuel cell (FC) device d3 are devices of a supplier
type. The electric company (EC) device d4 is a supplier type, which supplies
electric energy not only to the home (d0 ) device but also the RB (d1 ) device.
For each device di , a pair of a sensor si and actuator ai are installed. Each
device di is connected to an edge fog node fi as shown in Fig. 5. Each sensor si
gives electric power [w] which is supplied to other devices and is consumed every
one second. By each actuator ai , the electric power supplied and consumed by
a device di is controlled.
A node f0 is a root node which is a cloud servers. A fog node f0i exists for a
town (i = 0, ..., l0 − 1) and f0ij for each house (j = 0, ..., l0i − 1). Each edge fog
node f0ijk collects sensor data, i.e. electric power supplied and consumed five
devices d0 , d1 , d2 , d3 , d4 and sends the data to a parent fog node f0ij . Here, the
home device d0 just consumes electric energy supplied by the actuators a1 , ...,
a4 of the devices d0 , d1 , d2 , d3 , d4 . The battery (RB) device not only supplies
energy to the home device but also is supplied with energy from the solar panels
(SP), fuel cell (FC), and electric company (EC) devices.
Fig. 5. Example of TWTBFC model.
In this paper, we newly proposed the TWTBFC model to realize the IoT. In the
TBFC model, only a root fog node, i.e. server makes a decision on actions to
be executed by actuators. The actions are sent by the server to the actuators.
It takes time to deliver actions to actuators since edge nodes receive data from
sensors. In the TWTBFC model, fog nodes autonomously decide on actions by
using sensor data which are obtained by not only child fog nodes but also other
sibling fog nodes. Thus, each fog node can obtain more global data, i.e. data
from sibling nodes. We showed the electric power supply system as an example
of the TWTBFC model.
We are now evaluating the TWTBFC model.
620 Y. Guo et al.

15H0295.
References
1. Creeger, M.: Cloud computing: an overview. Queue 7(5), 3–4 (2009)
2. Enokido, T., Ailixier, A., Takizawa, M.: A model for reducing power consumption
in peer-to-peer systems. IEEE Syst. J. 4, 221–229 (2010)
3. Enokido, T., Ailixier, A., Takizawa, M.: An extended simple power consumption
model for selecting a server to perform computation type processes in digital ecosys-
tems. IEEE Trans. Ind. Inform. 10, 1627–1636 (2014)
4. Oma, R., Nakamura, S., Duolikun, D., Enokido, T., Takizawa, M.: An energy-
efficient model for fog computing in the internet of things (IoT). Internet Things
1–2, 14–26 (2018)
5. Oma, R., Nakamura, S., Duolikun, D., Enokido, T., Takizawa, M.: Evaluation of
an energy-efficient tree-based model of fog computing. In: Proceedings of the 21st
International Conference on Network-Based Information Systems (NBiS-2018), pp.
99–109 (2018)
6. Oma, R., Nakamura, S., Enokido, T., Takizawa, M.: A tree-based model of energy-
efficient fog computing systems in IoT. In: Proceedings of the 12th International
Conference on Complex, Intelligent, and Software Intensive Systems (CISIS-2018),
pp. 991–1001 (2018)
7. Rahmani, A.M., Liljeberg, P., Preden, J.-S., Jantsch, A.: Fog Computing in the
Internet of Things. Springer, Heidelberg (2018)
Security Intricacies in Cyber-Physical
Systems and Services
(INTRICATE-SEC-2019)
Stochastic Methods to Find Maximum
Likelihood for Spam E-mail Classification
Seyed M.-H. Mansourbeigi(&)
Department of Computer Science, College of Engineering, Utah State University,

4205 old Main Hill, Logan, UT 84322, USA
smansourbeigi@aggiemail.usu.edu
Abstract. The increasing volume of unsolicited bulk e-mails leads to the need
for reliable stochastic spam detection methods for the classification of the
received sequence of e-mails. When a sequence of emails is received by a
recipient during a time period, the spam filters have already classified them as
spam or not spam. Due to the dynamic nature of the spam, there might be emails
marked as not spam but are actually real spams and vice versa. For the sake of
security, it is important to be able to detect real spam emails. This paper utilizes
stochastic methods to refine the preliminary spam detection and to find maxi-
mum likelihood for spam e-mail classification. The method is based on the
Bayesian theorem, hidden Markov model (HMM), and the Viterbi algorithm.
1 Introduction
Spams are unwanted emails that the user does not want to have in his inbox. Email
Spam is one of the major problems of the today’s internet, bringing financial damage to
companies and annoying individual users. Spam is not only offensive and annoying; it
causes loss of productivity, decreases bandwidth and costs companies a lot of money.
Blocking spam email is considered a priority for network administrators and security
researchers. Spam filters are employed to assist the user in deciding if an email is worth
reading or not. There have been tremendous research efforts in this field that resulted in
a lot of commercial spam filtering products, such as: methods for construction of filters
to eliminate unwanted messages [1], comparison between the performances of machine
learning-based classifiers in filtering email spam [2, 3].
The challenging problem with spam filtering is the dynamic nature of the spam. The
characteristics (e.g. topics, frequent terms) of spam e-mail vary rapidly over time as
spammers always seek to invent new strategies to bypass spam filters [3]. One cannot
develop a filter and immediately implement it, because it will not have any basis for
classifying a document as spam or not spam. When a sequence of emails is received by
a recipient during a time period, the spam filters have already classified them as spam
or not spam. Due to the dynamic nature of the spam, there might be emails marked as
not spam but are actually real spams. It is important to be able to detect real spam
emails not only for personal convenience, but also for security. This paper proposes a
stochastic method to refine the preliminary spam detection. The method is based on the
Bayesian theorem, hidden Markov model (HMM), and the Viterbi algorithm.

https://doi.org/10.1007/978-3-030-15035-8_60
624 S. M.-H. Mansourbeigi
In the spam detection ground work, the Bayesian filtering works by evaluating the
probability of different words appearing in legitimate and spam mails and then clas-
sifying them based on those probabilities [4, 5]. A hidden Markov model (HMM) is a
simple dynamic Bayesian network that is characterized by the initial state, state tran-
sition, and emission probabilities. This statistical model is derived by assuming that the
system under consideration is a Markov process with hidden states. The HMM is able
to correct deliberate misspellings, incorrect segmentations (adding/removing spaces),
and other word modifications [6–8].
The approach in this paper answers the following questions:
(a) How to find the probabilities (transition, initial, and emission) for the Hidden
Markov model?
(b) What is the initial probability that a random email is spam or not spam?
(c) If the indication (from the emails’ header) for a finite sequence of emails on a
finite time period, is marked as spam or not spam, what is the legitimate sequence
of emails, based on the stochastic model.
2 Hidden Markov Model (HMM)
Given a sequence of states in different time frames (marked as spam and marked as not
spam), what are the most likely states that are congruent with the derived probabilities
and stochastic model based on training data. The training data in this case is the
confirmed and real emails identified as spam (S) or not spam (NS).
Suppose the training data is given by the sequence of emails:
S; S; S; S; NS; NS; S; S; S; S; NS; NS; S; S; S; ::; ::NS. . .; :. . .S. . .; . . .NS. . .; . . .S. . . ð1Þ
The transition probabilities are observed based on the number of times that the
switch from spam to not spam, not spam to spam happens and also the number of times
that the switch does not happen. These numbers are
S ! S 85 times NS ! S 35 times ð2Þ
S ! NS 15 times NS ! NS 65 times ð3Þ
Therefore the transition probabilities are shown in Fig. 1.

Based on the transition probabilities, the initial probabilities are calculated:
PðSÞ ¼ 0:85 PðSÞ þ 0:35 PðNSÞ PðNSÞ ¼ 0:15 PðSÞ þ 0:65 PðNSÞ ð4Þ
Since P(S) + P(NS) = 1, P(S) = 0.7 and P(NS) = 0.3. Meaning that without any
knowledge or assumption from the header of emails, P(NS) (the probability that an
email is not spam) = 0.3 (a priori).
Stochastic Methods to Find Maximum Likelihood 625
Fig. 1. Transition probabilities
The hidden states are marked as spam (MS) and marked as not spam (MNS). In
order to find emission probabilities, the Bayes’ theorem applies to conditional proba-
bilities calculation. The Bayes theorem states the prior probability (a priori): uncon-
ditional probabilities of our hypothesis before we get any data or any new evidence.
Simply speaking, it is the state of our knowledge before the data is observed. The
posteriori probability is a conditional probability about our hypothesis (our state of
knowledge) after we revised based on the new data [9].
For the hidden state MS, a posteriori, P(S|MS) = 0.85. Meaning that the probability
of the state to be spam is 0.85 if the hidden spam is marked as spam. For the hidden
state MNS, P(S|MNS) = 0.15. Meaning that if hidden state marked as not spam, then
the probability of the state to be spam is 0.15. Similarly, a posteriori, P(NS|MS) = 0.35
and P(NS|MNS) = 0.65.
Based on the training data, the emission probabilities emitted from the hidden states
are shown in Fig. 2.
Fig. 2. Emission probabilities
Based on all the probabilities the HMM is shown in Fig. 3.

Fig. 3. HMM
3 Maximum Likelihood Estimation (MLE) and Viterbi

Algorithm
The Viterbi algorithm is a dynamic programming algorithm for finding the most likely
sequence of hidden states, the Viterbi path that results in a sequence of observed events
especially in the context of hidden Markov models [10]. For a simple case in which
there are two emails with the hidden states at time period Time 1 and Time 2, given
respectively by MS and MNS, what are the actual states based on HMM?
For the two states S and NS there are 22 = 4 transitions:
S!S S ! NS NS ! S NS ! NS ð5Þ
Based on the HMM Fig. 3, the probabilities associated with the four transition
states are as in Fig. 4.
Therefore, MLE for this case is 0.076. Meaning that the actual transition states
based on HMM is S ! S for MS and MNS.
For the three emails with the hidden states at time period Time 1, Time 2, and Time
3, given respectively by MS, MNS, and MS, based on HMM, the actual states are
calculated as follows.
For the two states S and NS there are 23 = 8 transitions:
S ! S ! S NS ! S ! NS S ! NS ! S NS ! NS ! NS ð6Þ
NS ! S ! S S ! S ! NS NS ! NS ! S S ! NS ! NS ð7Þ
Based on the HMM, comparing the probabilities associated with the eight transition
states, the MLE for this case is MLE = 0.7 * 0.85 * 0.85 * 0.15 * 0.85 * 0.85 = 0.055.
Meaning that the actual transition states based on HMM is S ! S ! S, for MS, MNS,
and MS.
Fig. 4. Probabilities associated with the four transition states
With the same token for the case of a sequence with 4 time periods, there are
24 = 16 paths that are required to be checked for MLE. For 5 time periods, there are
25 = 32 paths and for 6 time periods there will be 26 = 64 paths. It is seen that the
number of paths increases exponentially.
In order to reduce the number of paths from 32, 64, and … to only one path, the
dynamical programming and specifically Viterbi algorithm is the natural approach to
achieve the solution.
Consider the sequence of hidden states MS, MS, MNS, MNS, MNS, MS given for
the time periods Time 1, 2, 3, 4, 5, 6 respectively. By utilizing the Viterbi algorithm,
based on HMM (Fig. 3), the MLE probabilities for the S and NS states are:
For the time period Time 1, between the two probabilities
PðSÞ ¼ 0:7 0:85 ¼ 0:595 PðNSÞ ¼ 0:3 0:35 ¼ 0:105 ð8Þ

The MLE is 0.595. Therefore the state S is the first candidate at Time 1.
For the time period Time 2, between the four probabilities
PðSÞ ¼ 0:595 0:85 0:85 ¼ 0:429 PðSÞ ¼ 0:105 0:35 0:85 ¼ 0:031 ð9Þ
PðNSÞ ¼ 0:595 0:15 0:35 ¼ 0:031 PðNSÞ ¼ 0:105 0:65 0:35 ¼ 0:024
ð10Þ
The MLE is 0.429. Therefore the state S is the next candidate at Time 2.
PðSÞ ¼ 0:429 0:85 0:15 ¼ 0:055 PðSÞ ¼ 0:031 0:35 0:15 ¼ 0:002 ð11Þ
PðNSÞ ¼ 0:429 0:15 0:65 ¼ 0:042 PðNSÞ ¼ 0:031 0:65 0:65 ¼ 0:013
ð12Þ
The MLE is 0.055. Therefore the state S is the next candidate at Time 3.
Up to here the sequence is S ! S ! S.
PðSÞ ¼ 0:055 0:85 0:15 ¼ 0:007 PðSÞ ¼ 0:042 0:35 0:15 ¼ 0:002
ð13Þ
PðNSÞ ¼ 0:055 0:15 0:65 ¼ 0:005 PðNSÞ ¼ 0:042 0:65 0:65 ¼ 0:018
ð14Þ
The MLE is 0.018. Therefore the state NS is the candidate at Time 4.

Therefore the sequence is extended by S ! S ! S ! NS.
PðSÞ ¼ 0:007 0:85 0:15 ¼ 0:0009 PðSÞ ¼ 0:018 0:35 0:15 ¼ 0:001
ð15Þ
PðNSÞ ¼ 0:007 0:15 0:65 ¼ 0:0007 PðNSÞ ¼ 0:018 0:65 0:65 ¼ 0:008
ð16Þ
The MLE is 0.008. Therefore the state NS is the candidate at Time 5.

Therefore the sequence of states is S ! S ! S ! NS ! NS.
Fig. 5. For the sequence MS, MS, MNS, MNS, MNS, MS the real states are:
S ! S ! S ! NS ! NS ! S
PðSÞ ¼ 0:001 0:85 0:85 ¼ 0:0007 PðSÞ ¼ 0:008 0:35 0:85 ¼ 0:0024
ð17Þ
PðNSÞ ¼ 0:001 0:15 0:35 ¼ 0:00005 PðNSÞ ¼ 0:008 0:65 0:35 ¼ 0:0018
ð18Þ
The MLE is 0.0024, the state S is the candidate at Time 6 (Fig. 5).
At this point the sequence is S ! S ! S ! NS ! NS ! S.
The concept of conditional probability and Bayesian theory is essential and pro-
found [11]
PðSjMSÞ ¼ PðMS AND SÞ=PðMSÞ ¼ ðPðMS j SÞ PðSÞÞ=PðMSÞ

¼ ðPðMS j SÞ PðSÞÞ=ððPðMSjSÞ PðSÞÞ ðPðMSj SÞ Pð SÞÞÞ
ð19Þ
4 The Python Program for Viterbi Algorithm
# solving with dynamical programming and viterbi algorithm

# Transition Probabilities
# S = spam NS = not spam
# MS = marked as a spam MNS = marked as not spam -- mean
/feelguess it is spam or feel/guess it is not spam
p_ss = 0.85 # p marked as a spam goes to p marked as a spam
- probability for next time around
p_sr = 0.15 # p marked as a spam goes to p marked as a not
spam - probability for next time around
p_rs = 0.35 # p marked as a not spam goes to p marked as a
spam - probability for next time around
p_rr = 0.65 # p marked as a not spam goes to p marked as a
not spam - probability for next time around
# Initial Probabilities
p_s = (p_rs)/(p_rs + p_sr)
p_r = 1 - p_s
#print(p_s)
#print(p_r)
#p_s = 0.7 # p to be marked as a spam - overall probability
#p_r = 0.3 # p to be marked not as a spam - overall
probability
# Emission Probabilities
p_sh = 0.85 # p to be marked as a spam and turns a spam -
probability from hidden states
p_sg = 0.15 # p to be marked as a spam and turns not a
spam - probability from hidden states
p_rh = 0.35 # p to be marked not as a spam and turns a
p_rg = 0.65 # p to be marked not as a spam and turns not a
finals = ['S' , 'NS' , 'NS' , 'S' , 'NS', 'S' , 'S' , 'S' ,
'NS' , 'S' , 'NS']
probabilities = []
primary = []
if finals[0] == 'S':
probabilities.append((p_s*p_sh, p_r*p_rh))
else:
probabilities.append((p_s*p_sg, p_r*p_rg))
for i in range(1,len(finals)):
yesterday_sunny, yesterday_rainy = probabilities[-1]
if finals[i] == 'A':
today_sunny = max(yesterday_sunny*p_ss*p_sh,
yesterday_rainy*p_rs*p_sh)
today_rainy = max(yesterday_sunny*p_sr*p_rh,
yesterday_rainy*p_rr*p_rh)
probabilities.append((today_sunny, today_rainy))
else:
today_sunny = max(yesterday_sunny*p_ss*p_sg,
yesterday_rainy*p_rs*p_sg)
today_rainy = max(yesterday_sunny*p_sr*p_rg,
yesterday_rainy*p_rr*p_rg)
probabilities.append((today_sunny, today_rainy))
print(probabilities)
for p in probabilities:
if p[0] > p[1]:
primary.append('MS')
else:
primary.append('MNS')
print (primary)
print(finals)
For the marked sequence [‘MS’, ‘MS’, ‘MNS’, ‘MS’, ‘MNS’, ‘MS’, ‘MS’, ‘MS’,
‘MS’, ‘MS’, ‘MS’] as an input, by executing the above python program, the output is:
[‘S’, ‘NS’, ‘NS’, ‘S’, ‘NS’, ‘S’, ‘S’, ‘S’, ‘NS’, ‘S’, ‘NS’].
The method in this paper can be utilized to address the following future works:
1 – Speech recognition
2 – Genetics DNA sequences
3 – Vehicle location
4 – Text tagging WORDS to TAGS (the definite articles, preposition, verb, noun).
References
1. Sahami, M., Dumais, S., Heckerman, D., Horvitz, E.: A Bayesian approach to filtering junk
e-mail. In: AAAI 1998 Workshop on Learning for Text Categorization (1998)
2. Al-Jarrah, O., Khaterz, I., Al-Duwairi, B.: Identifying potentially useful email header
features for email spam filtering. In: ICDS The Sixth International Conference on Digital
Society (2012)
3. Awad, W.A., ELseuofi, S.M.: Machine learning methods for spam e-mail classification. Int.
J. Comput. Sci. Inform. Technol. 3(1), 173–184 (2011)
4. Eberhardt, J.J.: Bayesian spam detection. Scholarly Horizons: University of Minnesota,
Morris Journal, vol. 2, no. 1 (2015)
5. Freeman, D.M.: Using naive Bayes to detect spammy names in social networks. In:
Proceedings of the 2013 ACM Workshop on Artificial Intelligence and Security, AISec
2013, pp. 3–12. New York (2013)
6. Lee, S., Jeong, I., Choi, S.: Dynamically weighted hidden Markov model for spam
deobfuscation. In: IJCAI 2007, Proceedings of the 20th International Joint Conference on
Artificial Intelligence, pp. 2523–2529 (2007)
7. Lee, H., Ng, A.Y.: Spam deobfuscation using a hidden Markov model. In: Proceedings of
the 2nd Conference on Email and Anti-Spam, Stanford University (2005)
8. Jaswal, V., Sood, N.: Spam detection system using hidden Markov model. Int. J. Adv. Res.
Comput. Sci. Softw. Eng. 3(7), 304–308 (2013)
9. Roy, S., Patra, A., Sau, S., Mandal, K., Kunar, S.: An efficient spam filtering techniques for
email account. Am. J. Eng. Res. 2(10), 63–73 (2013)
10. Viterbi, A.J.: Error bounds for convolutional codes and an asymptotically optimum decoding
algorithm. IEEE Trans. Inf. Theor. 13(2), 260–267 (1967)
11. Papoulis, A.: Probability, Random Variables, and Stochastic Processes. McGraw-Hill Series
in Electrical Engineering, 3rd edn. (1991)
Design and Research of Trusted Acquisition
Terminals Based on Domestic Password
Yun Feng(&), Xiaobing Liang, Feng Zhai, and Meng Xu
China Electric Power Research Institute, Beijing, China

{fengyun,liangxiaobing,zhaifeng,
xumeng}@epri.sgcc.com.cn
Abstract. As one of the core equipment of power consumption information

collection system, the acquisition terminal carries a large amount of data and
faces many information security risks. This paper analyzes the trusted com-
puting system based on the China’s own cryptographic algorithm, combs the
security risks faced by the acquisition terminal, and applies the trusted tech-
nology based on the China’s own cryptographic algorithm to the security
reinforcement of the acquisition terminal. The design and research of the trusted
acquisition terminal based on TCM is carried out from three aspects: the overall
architecture, trusted start and trust chain transmission, which can improve the
security and protection ability of the acquisition terminal effectively.
1 Introduction
With the gradual development of smart grid in China, the construction of power
consumption information collection system is becoming more and more perfect. Fre-
quent data interaction of power consumption information collection system brings
security risks [1]. According to statistics, most of the security attacks are initiated by
terminal devices, which are vulnerable to various types of security threats and attacks,
such as denial of service attacks, eavesdropping, illegal intrusions, and so on. These
risks will lead to the leakage of sensitive information and the control of the terminal by
hackers, thus threatening the normal operation of the whole system [2, 3]. Although it
is one of the security weaknesses in the system, the security of the terminal system has
not received enough attention.
The concept of trusted computing makes people transfer the idea of solving
information security problems to solving terminal security problems [4]. At present, the
research in the field of trusted computing in China has gone from the initial tracking
learning to the stage of independent innovation. At the same time, China attaches great
importance to information security, and chinese codes have been developing rapidly.
The application of domestic codes in trusted computing can effectively reverses the
current situation that the security protection technology of key information systems in
our country relies heavily on foreign countries [5], which can improve the autonomous
and controllable ability in the field of security technology in our country.
China Electric Power Research Institute Youth fund (JL84-18-001).

https://doi.org/10.1007/978-3-030-15035-8_61
634 Y. Feng et al.
Therefore, this paper combines the current status of power consumption informa-
tion collection system, applies the trusted computing technology based on domestic
cryptographic algorithm to the security reinforcement of terminal equipment, and puts
forward the security design scheme of the acquisition terminal, which can effectively
improve the security of the acquisition terminal, and have urgent realistic needs and
research significance.
2 The Trusted Computing System Based on Domestic

Cryptography Algorithm
2.1 The Overview of Trusted Computing
The concept of trusted system was first proposed by Anderson JP in early 1970s. In
1999, IBM, HP, Intel and Microsoft jointly established Trusted Computing Platform
Alliance (TCPA), which was reorganized into Trusted Computing Group (TCG) in
2003. As a non-profit organization, TCG aims to promote the development of trusted
computing platform standards. At present, it has formulated a series of trusted com-
puting technical specifications [6–9].
Trusted computing technology starts with the trusted roots guaranteed by physical,
technical and management security, and establishes a trust chain which start from
hardware platform, operating system, to upper application. Based on the trust root, it
can authenticate and trust stair from one to another, and expand this kind of trust to the
whole trusted compute field, so as to ensure the trustworthiness of the whole system
environment [10].
2.2 TCM Based on the Domestic Password

Cryptography plays an important role in trusted computing. Trusted computing
involves not only cryptographic algorithms, but also cryptographic protocols, key
management, certificate configuration and many other aspects. China’s domestic
ciphers have developed rapidly in recent years. Many domestic cipher algorithms, such
as SM1, SM2, SM3, SM7, have been proposed. They have been applied in security
protection in power, finance, Internet and other fields.
The core component of trusted computing is trusted computing platform module
(TPM) [11, 12]. In December 2007, CTCP, China Trusted Computing Platform Alli-
ance, formulated its own Trusted Cryptography Module (TCM) specification. Con-
sidering China’s specific national conditions, the specification has made many
improvements over TCG’s TPM standard. It is mainly embodied in three aspects: based
on domestic cryptographic algorithm, using symmetric and asymmetric cryptography,
and using double certificate system. The basic structure of TCM is shown in Fig. 1.
In terms of algorithm usage, passwords used in TCM are all domestic. The hash
algorithm is changed from SHA1 to SCH; the asymmetric cryptographic algorithm uses
elliptic curve digital signature algorithm (SM2-1), elliptic curve key exchange protocol
(SM2-2), elliptic curve public key encryption algorithm (SM2-3); besides, TCM also
Design and Research of Trusted Acquisition Terminals 635
SM2 Random Execution Volatile

Number
Engine Generator Engine Memorizer
I/O
Port
SMS4 SM3 HMAC Nonvolatile

Engine Engine Engine Memorizer
Fig. 1. TCM component structure.
adds symmetric cryptographic algorithm, which reduces the complexity of key man-
agement, data encryption and authorization protocol. The SMS4 is used as a symmetric
cipher algorithm.
3 A Security Reinforcement Method of Trusted Acquisition

Terminals Based on Domestic Password
3.1 The Security Risk Analysis of Acquisition Terminals

As the key equipment in power consumption information collection system, the
acquisition terminal plays a connecting role in the system: communicating with the
main station system to achieve remote data transmission; communicating with smart
meters to collect and process meter data, reducing the burden of data storage and
processing. Because of its special deployment location and function, the acquisition
terminal faces many security risks, as shown in Table 1.
Table 1. Security risks faced by acquisition terminals.

Risk type Risk description
Physical attack The attacker directly interrupts the acquisition
terminal itself or the link line for physical damage,
resulting in the leakage of electricity data and users’
privacy
Information is eavesdropped, forged, By illegal means, the attacker wiretaps, forges,
replayed and tampered with replays and tampers the data in the acquisition
terminal, which brings security risks to the normal
operation of the electric power information
acquisition system and the privacy of user data
Denial of service attack The attacker consumes the network bandwidth of the
acquisition terminal illegally, which makes the
acquisition terminal unable to continue to work
In view of various security threats, the current security reinforcement measures

adopted by the acquisition terminal include physical packaging, circuit protection,
security hardware and so on. However, due to the complex deployment environment of
636 Y. Feng et al.
the acquisition terminal, the acquisition terminal can not be effectively monitored,
physical protection can not effectively prevent attacks; there is still a certain gap
between the security hardware encryption capability and the data requirements of the
acquisition terminal; for the operation system and application program which under-
takes the main work of the acquisition terminal, there is still no good detection means;
the acquisition terminal currently adopts modular design, even if the functional mod-
ules are safe, there will still be security risks in the combination of the modules [13].
3.2 The Working Principle of Trusted Acquisition Terminals Based

on TCM
As one of the most important terminals to realize “full coverage, full acquisition, full
cost control” in China’s power consumption information collection system, the safety
protection research of acquisition terminal is very necessary. Trusted computing pro-
vides a new idea for the security protection of the acquisition terminal, which can
fundamentally improve the security of the acquisition terminal.
The basic idea of building a trusted acquisition terminal is to implant TCM chip
into the acquisition terminal system. By calling the protection function, integrity
measurement function and authentication function of TCM, it can realize the security
and credibility of the acquisition terminal system effectively.
The key to the trusted collection terminal system based on TCM is to realize trusted
startup and pass the trust step by step. TCM-based trusted acquisition terminal start-up
should start with TCM trusted root, TCM needs to verify the integrity of Bootloader
(operating system loader), if success, Bootloader can be started and get control rights;
Bootloader then verifies the integrity of the operating system by calling the function of
TCM. After the operation system is trusted, the operating system starts running and
gets control from Bootloader. After that, the operating system verifies the trustwor-
thiness of the application through the trustworthiness measure, and transfers control to
the application. Furthermore, the application program verifies the security and trust-
worthiness of each application step by step, and transfers the control rights in turn to
realize various trustworthy applications through step-by-step measurement and step-by-
step transfer.
The trusted acquisition terminal system guarantees the trusted start-up and opera-
tion of the underlying hardware, operating system and upper application program
through trusted step-by-step transmission, and enhances the security protection
capability.
4 The Design of Trusted Acquisition Terminals Based

on TCM
4.1 The General Architecture of Trusted Acquisition Terminals

The overall architecture of the trusted acquisition terminals is shown in Fig. 2.
The difference between the trusted acquisition terminal and the general acquisition
terminal is that TCM trusted chip is added. TCM in the trusted chip acts as the trusted root
Fig. 2. General architecture of trusted acquisition terminals.
and cooperates with the trusted acquisition terminal operating system to realize the
integrity, identity trustworthiness and data security protection of the acquisition terminal.
Trusted acquisition terminal architecture mainly includes three parts: trusted
hardware system, trusted operating system and trusted application program. In the
hardware system of the trusted acquisition terminal, TCM communicates with the main
control unit module of the acquisition terminal through the trusted module interface
circuit (TMIC). TMIC also has the function of assisting TCM to complete the trusted
verification and start-up control of Bootloader, which will be introduced in detail later.
Trusted operating system mainly includes software layer and service layer. Based on
TCM supporting software TDDL, TCS and TSP, it completes three functions: trusted
measurement, trusted message and trusted storage, realizes integrity verification,
identity authentication and security protection of acquisition terminal, and provides
trusted service for acquisition terminal application.
4.2 The Trusted Startup of Trusted Acquisition Terminals

The module structure of the terminal main control unit of the trusted acquisition system
is shown in Fig. 3. It is mainly composed of CPU basic circuit, input and output circuit,
communication interface circuit, speech control circuit, trusted module interface circuit
and other circuits. The trusted module interface circuit is unique to the trusted acqui-
sition terminal, which is used to communicate with the trusted cryptographic module
TCM chip. The basic circuit of CPU includes microprocessor MCU, data memory,
program memory, clock circuit, auxiliary circuit and so on.
638 Y. Feng et al.
MCU
ARM7
Input Communication
/Output Interface Circuit
Data Memory
Circuit
Program
Memory
Trusted Module
Interface Circuit
Clock Circuit
Auxiliary
Other Circuit Speech Control
Circuits Circuit
CPU
Basic Circuit
Fig. 3. Block diagram of main control unit module of acquisition system
Trusted module interface circuit (TMIC) is the key to the realization of trusted
acquisition terminal system. Besides providing communication channel between TCM
and MCU, it also has the function of assisting TCM to complete trusted boot loader.
The specific process of trusted start-up of trusted acquisition terminal system is as
follows:
(1) TCM starts and checks itself first after the system is powered up;
(2) TMIC reads Bootloader to TCM and verify its integrity;
(3) If the integrity checks pass, TMIC establishes the actual connection between
MCU and Bootloader, TCM sends license signal to MCU processor, and Boot-
loader starts; if the check is wrong, reset the system.
4.3 The Trust Chain Transfer of Trusted Collection Terminals

The key of trust chain transfer in trusted acquisition terminal system is how to start
trustfully and realize trust transfer step by step [14, 15]. After the completion of trusted
startup, the operating system starts to run and transfers trust to the application program.
Because of the numerous application systems, the transfer of trust chain can take many
forms, such as tree, depending on the specific situation. The trust chain transfer of
trusted collection terminals is shown in Fig. 4.
The trust chain transfer process of trusted acquisition terminals is as follows:
Step 1. system power up, TCM initialization, start self-testing;
Step 2. TCM reads Bootloader to TCM through TMIC.
Step 3. TCM calculates the hash value of Bootloader and compares it with the
existing benchmark value to check the integrity of Bootloader. If Bootloader passes
through, enters step 4 and TCM will give control to Bootloader, otherwise it will
return to step 1.
Step 4. Bootloader establishes the actual connection with MCU through TMIC.
TCM transmits license signal to MCU through TMIC to start Bootloader.
Turn on the system
TCM self-check
TMIC error
bootloader
check
Success TMIC
MCU
Bootloader start
operating system
Application
Fig. 4. Trust chain transfer of trusted acquisition terminals
Step 5. Bootloader calls TCM’s interface command, calculates the hash value of
operating system, and completes the integrity check by comparing with the baseline
value. If the check passes, Bootloader transfers control to the operating system and
enters step 6, otherwise it restores and reports errors.
Step 6. The operating system calculates the hash value of the application and
completes the integrity check by comparing with the baseline value. After the check
is passed, the control will be given to the application and enter step 7. Otherwise,
the application will be restored and the error will be reported.
Step 7. application goes into operation. By measuring step by step and transferring
trust chain in such a trusted way, we can verify the security and trustworthiness of
each application, and fundamentally protect the security of the whole collection
terminal system.
5 The Concluding Remarks
The gradual informationization and interaction of smart grid make the security of
power information acquisition terminal more and more important. In recent years, the
domestic cryptography technology and trusted technology have developed rapidly in
our country. In this paper, the trusted computing technology based on the domestic
cryptography is applied to the security reinforcement of the terminal of power infor-
mation acquisition system. The design of trusted acquisition terminal is studied from
three aspects: overall architecture, trusted start-up and trust chain transmission, which
640 Y. Feng et al.
can effectively enhance the safety protection ability of acquisition terminals and bring
new ideas for the safety reinforcement of the acquisition terminals.
References
1. Zhang, S.M., Wang, Z.Y., Wang, B.Y.: Terminal integrity detection scheme of electricity
information acquisition system based on trusted computing. Electr. Power Autom. Equip. 37
(12), 60–66 (2017)
2. The Research and Implementation of Embedded Power Distribution Terminal Based on
Trusted Computing. North China Electric Power University, Beijing
3. Shen, C.X., Zhanu, H.U., Fend, D.U., et al.: Survey on information security. Sci. China Ser.
E Inf. Sci. 37(2), 129–150 (2007)
4. Department of Defense Computer Security Center: Department of Defense Trusted
Computer System Evaluation Criteria. DOD 5200.28-STD. DOD, USA (1985)
5. Shen, C.X., Gong, B.: The innovation of trusted computing based on the domestic
cryptography. J. Cryptol. Res. 2(5), 381–389 (2015)
6. Trusted Computing Group. TCG [OL]. https://www.trusted.computinggroup.org
7. Trusted Computing Group [OL]. https://www.trusted.computinggroup.org/specs/
8. Trusted Computing Group, TCG Specification Architecture Overview [EB/OL]. Revision
1.4 (2007)
9. Trusted Mobile Platform Protocol Specification Document [EB/OL]. http://www.trusted-
mobile.org/TMP_Protoeol_revl_00.Pdf
10. Yang, B., Wu, Z.-Q., Fu, X.-P.: Dynamic integrity measurement model based on trusted
computing. Comput. Eng. 38(2) (2012)
11. Yu, Z., Zhang, W., Dai, H., et al.: A trusted architecture for virtual servers with trusted
platform module and certify-cafe authority. J. Signal Process. Syst. 86(2/3), 327–336 (2017)
12. Mukhamedov, A., Gordon, A.D., Ryan, M.: Towards a verified reference implementation of
a trusted platform module. In: Security Protocols XVII, pp. 69–81. Springer, Heidelberg
(2013)
13. Zhai, B., Ni, M., Shi, Y., Fan, P.: A review on the security of embedded system. J. Wuhan
Univ. (Nat. Sci. Ed.) 64(2), 95–108 (2018)
14. Xu, M.D., Zhang, H.G., Zhang, F., et al.: Survey on chain of trust of trusted system. Acta
Electron. Sin. 42(10), 2024–2031 (2014)
15. Chang, D.X., Feng, D.G., Qin, Y., et al.: Analyzing the trust chain of trusted virtualization
platform based on the extended LS2. J. Commun. 34(5), 31–41 (2013)
Trust-Based Security Mechanism for Detecting
Clusters of Fake Users in Social Networks
Davinder Kaur, Suleyman Uslu, and Arjan Durresi(&)
Department of Computer and Information Science,

Indiana University Purdue University, Indianapolis, IN, USA
{davikaur,suslu}@iu.edu, adurresi@iupui.edu
Abstract. Inspired by the ability of the trust mechanism which can capture
human trust using measurement theory we have designed trust-based security
mechanism to find clusters of fake users in social networks. Proposed framework
enables us to take human and community knowledge into a loop by using a
variation of trust to detect fake users. This framework will increase the accuracy
of detecting fake users by using both human and machine knowledge. Experi-
ments are performed on randomly generated social networks to validate the
potential of this framework. Results show that the variation of trust over time
can able to differentiate between clusters of fake users from clusters of real users.
1 Introduction
With the rise in the use of technology, the use of social media is also increasing. Lots of
people use social media to share pictures, videos, news and to stay up to date with the
current events and news. People are also using social networks to connect with other
people. However, with so many advantages of social media, there are some disad-
vantages also. Bad people are misusing social media to spread fake news, malicious
content and viruses by creating fake user accounts, that cannot easily be caught. So, it
is imperative to detect these fake user accounts as soon as possible to minimize the
harm caused by them.
To detect fake user accounts, it is essential to understand how social networks
work. All our social interactions whether they are face to face or over the Internet are
based on the trust which is based on the number of positive and negative interactions
happened between the people and is developed over time and can be used to detect fake
users. [1] In this paper a trust-based mechanism is proposed for social networks to filter
out malicious content and to study the damage caused by the users based on their trust
values. This type of mechanism is helpful in studying the behavior of fake users.
We have developed a method to detect fake accounts on the social network based
on a trust-based security mechanism. In this method, we have integrated trust-based
mechanism with machine learning to detect clusters of fake accounts. The trust-based
mechanism is used to capture human interactions with each other, and machine
learning is used to process and analyze a large amount of social network data. Trust-
based mechanism is used to map human interactions into physical measurements. This
mechanism will help to capture the human knowledge about other humans or events

https://doi.org/10.1007/978-3-030-15035-8_62
642 D. Kaur et al.
which can be very beneficial for detecting fake users. Machine learning algorithms are
used to capture characteristics of the social networks. It is used to cluster the social
network graph based on their characteristics. These clusters are then analyzed based on
the trust values, density, the variation of the trust values over time to label them as a
cluster of legitimate users or cluster of fake users. The proposed method of detection
can be beneficial as it can capture the community information to detect fake users.
This paper is organized as follows. In Sect. 2 we present the background and
related work in the fields of fake user detection and trust mechanism. Section 3
describes our framework of Fake user detection. In Sect. 4 we present different
experiments and simulations results of our framework. Also, in Sect. 5, we conclude
the paper.
This section gives information about the background and related work done in the field
of trust and security.
2.1 Fake Accounts Detection in Online Social Networks

One of the significant issues faced by social media is the detection and elimination of
fake accounts. Users create fake accounts to perform malicious activities such as
spreading false news, viruses and spamming.
Much research is done in this field for the early detection of fake users. Some
researchers used social network graph properties to detect fake users while others used
machine learning algorithms on the different feature space to detect fake profiles. Cao
et al. [8] analyzed social graph properties by performing random walks on the graph to
rank users based on the probability of being fake. Some researchers used supervised
machine learning to build classifiers for the detection of fake users [9, 10]. Xiao et al.
[12] used user profile information for building the classifier to detect clusters of fake
users. Researchers also used the similarity matrix to compute the similarity between the
actual user profile and fake profiles using different clustering methods [13]. All these
methods have their advantages and disadvantages. Some of these methods are com-
putationally less efficient than others, and some have higher accuracy than others. All
these researches in this field indicate how important it is to detect and stop fake users
from spreading false content.
Fake users are more determined in spreading false information because of the large
incentives they get by doing so. Fake users try to camouflage real users identity by
making more friends or by imitating real users so that they cannot be detected by fake
user detection techniques [11]. So, there is a need to build methods that consider the
features of user profiles, social network graph and most importantly human and
community knowledge about users to detect false profiles. Using these methods, the
local community can able to see the user profile and rate it according to their intuition
of profile being real or fake. Using these ratings by the users, we can calculate trust
between users that can be very helpful in detecting fake users.
Trust-Based Security Mechanism for Detecting Clusters of Fake Users 643
There are several approaches where the community information is used in detecting
fake users [1, 5, 8]. Some of the approaches used security experts to manually check
user profiles which the system ranked them as more likely to be fake to keep human
knowledge in the loop for detecting fake users [14]. Some of the approaches used trust
framework to calculate the damage caused by attackers based on their trust value
assigned by the local community [1].
In our approach, we have used the trust mechanism to capture local and community
knowledge used to detect fake clusters.
2.2 Trust Mechanism

In social networks, every relation between users is based on trust. There is a need for
some computational framework that can capture that relationship. Several works are
done regarding this field, one of which is using Trust based framework [5] for inferring
trustworthiness of social relations. This framework is used to model a trust-based
relationship between users or entities into computational form and infer indirect rela-
tionships. This technique of trust is very useful in social networks as all our social
relations are formed on trust, which is very difficult to capture when there is huge
number of connections [6].
The main advantage of the trust framework proposed by Ruan et al. [5] is that it is
flexible enough to be used in a vast variety of applications. The applications range from
using a trust for the security of IoT devices [3] to stock market prediction using social
network data [4].
There are also several examples of using the trust framework in online social
communities [2, 6, 7]. Ping [1] used the trust framework for information filtering in
social networks. Using the trust mechanism, he calculated the trustworthiness of the
sender who is sending the information to detect if the information sent by him is fake or
not. [8] used trust propagation via the power iteration method for detecting attackers in
online social communities.
All the above applications of trust framework validate its potential to be used in
detecting fake users in social networks. Using the trust framework in detecting fake
users will bring the human knowledge in the loop along with the automated system.
3 Fake User Detection Using Trust-Based Security

Mechanism
To make fake user detection more efficient, we have introduced Trust-based Security
Mechanism. This approach considers human and community knowledge along with
machine learning to detect clusters of fake users. We have designed evaluation metrics
based on trust between the nodes of the clusters and using features of social network
graph to label detected clusters as the clusters of fake users or the clusters of good
users.
This algorithm consists of three steps. The first step is to calculate the trust of the
users of the social network using their interactions with each other. Secondly is to
644 D. Kaur et al.
detect clusters in social network graph using a machine learning algorithm. Last is to
analyze clusters based on trust to label them as good or bad clusters.
Section 3.1 presents Trust framework used to calculate trust between the users of the
social network. Section 3.2 explains the clustering algorithm we have used to cluster
social network graph Sect. 3.3 explains the evaluation metrics used to label the clusters.
3.1 Trust Framework Inspired by Measurement Theory

Social interactions between the humans are based on trust, and this trust between
humans do not develop overnight, it is a long process. Trust is developed based on the
interactions and experiences between the people. If two people have more good
experiences with each other, then their trust value will be higher than the two people
who have a bad experience with each other. Human trust is based on the experience,
which can be used to predict the future experience. Trust mechanism helps to map
human trust to physical measurements [5]. This mechanism helps to quantify human
trust into measurement theory.
In social networks, trust between users is calculated based on the communications
happened between them. Different social networks have different attributes used to
calculate trust. In Facebook, likes and comments can be used as the measurement of
trust between users while in twitter retweets, mentions, reposts can be used as the
measurement of trust [4]. Trust framework used to calculate trust consists of two
metrics: Impression and Confidence.
Impression(m): Impression is defined as how much one person trusts another
person. That is, it is the summary of the trustworthiness of one person calculated by
another person based on his/her experiences with him. The experiences or interactions
can be direct or indirect.
In social networks each interaction act as a measurement between the truster and
trustee. The impression in social networks is calculated as the mean of all the mea-
surements taken between the truster and trustee as shown in Eq. 1.
Pn
mi
m¼ i¼1
ð1Þ
n
Confidence(c): Confidence is defined as how much a person is confident about

his/her impression(m) about another person. That is, how sure is the person is about the
impression of another person. This metrics is used to reduce the errors that occur in
calculating the impression. In measurement theory, confidence is related to the vari-
ance. If a person has high confidence about his impression about another person, that
means there is less variance in measurements. Equation 2 shows how confidence is
calculated using variance.
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
sP
n 2
i¼1 ðmi mÞ
c ¼ 1 2 v where v ¼ ð2Þ
nð n 1Þ
Both impressions and confidence are used to calculate the value of trust. We
normalized the values of impression and confidence between 0 and 1, but it can be
varied based on the circumstances. These values are used to represent trust between
users of the social network.
However, if there are no interactions happened between the nodes, then trust is
calculated using indirect links. Trust framework has two methods to calculate indirect
trust between the nodes, that are Trust Transitivity and Trust Aggregation.
Trust Transitivity: This method is used to calculate indirect trust. If a person X does
not have a direct connection to Z, but X knows someone for example Y who knows Z,
then trust from X to Z is calculated through Y using trust transitivity. So, the trust
between X and Z will be concatenation between trust from X to Y and Y to Z.
Equation 3 and four show how impression and confidence are calculated for transitive
trust.
Z ¼m
mXY mYZ ð3Þ
XY
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Z ¼
vXY ðvXY Þ2 ðmYZ Þ2 þ ðvYZ Þ2 ðmXY Þ2 ð4Þ
Trust Aggregation: This method is used to summarize trust value if there are more
than one indirect path between nodes. For example, to reach D, there are two paths A-
B-D and A-C-D from A then the aggregated trust will be the combination of both the
trust values of the graph. Equation 5 and six show how impression and confidence are
calculated for trust aggregation.
mADB þ mC
AD
B mC ¼
mAD ð5Þ
AD
2
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2 AD 2
ðvADB Þ þ vC
B vC
vAD ¼ ð6Þ
AD
22
Using this trust framework, we can calculate the trust of every user based on his/her
interactions with other users on social networks.
3.2 Clustering Social Network Graph

Machine learning mechanism is used to process and analyze massive amount of social
network data efficiently. Machine learning helped us to cluster large social network
graphs based on the similarity matrix between users. Using clustering we divide the
whole network into small clusters with similar properties, and on these clusters,
evaluation metrics based on trust is applied to detect clusters of fake users.
Today’s social network can be modeled into undirected graph G (V, E) composed
of nodes and edges. Nodes(V) corresponds to users of social network and edges(E)
corresponds to the relationship between the users. Trust is the weight of the edges
which is calculated based on the trust mechanism explained in the previous section.
646 D. Kaur et al.
To divide the social network graph into clusters, the spectral clustering method is
used. This clustering method has excellent performance in graph-based clustering and
is a combination of k-means clustering and DBSCAN.
The main aim of clustering the social network graph is to divide the graph into
groups of highly correlated subgraphs that share similar properties. That is, to maxi-
mize intracluster similarity and minimize intercluster similarity. This is performed to
deal with a large amount of social network data, as it would be more efficient to apply
the trust mechanism on graphs with similar properties.
After these clusters are detected, they are analyzed based on trust mechanism and
other features to label them as the clusters of good users or the cluster of fake users.
Machine learning algorithms enable us to cluster a large amount of data efficiently.
This makes the process of detecting fake users very fast. In the following section, we
will discuss the evaluation metrics based on which fake users are detected.
3.3 Evaluation Metrics

After detecting the clusters, it is vital to study their properties to categorize them as a
cluster of legitimate users or cluster of fake users. To develop the evaluation metrics,
we have assumed that fake users tend to establish more links between themselves and
very few with real users and the trust propagation of these connections are extremely
different from the connections of real users in the social network. Based on these
assumptions following are the characteristics which we have used to evaluate the
clusters:
The density of the Clusters: Density of clusters is used as one of the properties in
detecting fake clusters. Clusters of real users are more spread out then the clusters of
fake users as fake users tend to make more connections within themselves and fewer
connections to the outside world. If they try to make connections with real users, it is
very rare that real users will accept it or if they accept it, they will accept it with very
low trust value. This characteristic of fake users allows us to use density as one of the
characteristics to label cluster of fake users. The density of a graph is defined as the
ratio of some edges present in the graph to the maximal number of edges that graph can
have.
2 Number of edges in the graph

Density ¼
Number of VerticesðNumber of Vertices 1Þ
There can be a case where the density of the cluster of real users is also high, in that
case, the trust of the cluster and variation of trust over time makes the difference. So,
the density of the cluster is used along with trust to detect fake users.
Cluster formation based on time: Time taken by the fake users to build connections
within themselves is very less than the time taken by the real users to build actual
connections. Both the density of the cluster and time taken to form the clusters are
interrelated. Clusters or communities formed by fake users are within themselves, and
fake user profiles are only created if they need to perform some malicious activity, so
these clusters are formed instantly while the clusters of the real users take a significant
amount of time to stabilize and form communities. These features imply that clusters
formed at the instant of time are more probable of being clusters of fake users then the
clusters formed in a significant amount of time. Using this feature of time along with
trust feature enable us to detect clusters of fake users.
Variation of trust over time: Trust is the essential feature to detect fake users. Trust
is used to capture local information about the users. Trust-based metrics are used to
differentiate real users from fake users. These metrics are divided into two parts: First is
based on the average trust of the clusters and Second is based on the variation of trust
over time.
Fake user clusters tend to have a high trust value as compared to real user clusters.
This is based on how fake users make connections with other fake users. For spreading
malicious content, fake users need to increase their trustworthiness among other users
because if their trust is low, they can easily be detected as fake users. They do so by
exchanging positive messages with other fake users to increase their trust instantly.
Trust of fake user tend to have fewer fluctuations, and it will mostly increase with
time, while the trust of real users will have move fluctuations and it will go more up
and down based of the positive and negative interactions happened between them.
So, by analyzing the trust of the clusters and how the trust changes with time along
with the properties of density and time, fake user clusters can be detected.
4 Implementation
This section explains the implementation of our method.
4.1 Dataset
As the fake user data of social networks is not available publicly, we generated syn-
thetic social network graph that can mimic the properties of the real graph to experi-
ment our approach.
To simulate the social network graph, we have used the Erdos-Renyi model to
generate the random social graph. This type of model is very useful in generating
random social graphs [15]. If G (V, E) is the random graph generated, then V represents
users of the social network and E represents the relationship between users, which is
chosen randomly with the probability p such that p belongs to set [0, 1]. The degree of
these nodes is also random numbers selected from the binomial distribution [15].
Because of the randomness feature of this graph model, it is perfect for generating
random social graphs.
We have created a social graph of 1000 nodes, with less probability of edge
creation. This created graph mimics the social network graph of good users as the
nodes in this graph is distributed and sparsely connected. To mimic fake user behavior,
we increased the probability of edge creation so that the created graph is more densely
connected. Timestamp at which edge is created is also randomly generated.
To capture the trust values between the nodes, values of impression and confidence
are randomly generated. This generation of trust is divided into two sections. In one
section trust values is randomly assigned between [0, 1] for sparsely connected
648 D. Kaur et al.
distributed graph and for densely connected graph higher trust values are assigned
between [0.5, 1] to mimic the behavior of fake users.
All the data generated by the above methods are used to test our approach of
detecting fake user clusters.
4.2 Test and Results

Once the data is generated, we have performed experiments to test our approach. To
detect the clusters of fake users, the foremost thing is to detect clusters and then analyze
the detected clusters based on trust to detect clusters of fake users.
For clustering, we have used a machine learning algorithm to detect clusters. We
have experimented by detecting a different number of clusters. In one case we have
detected 2 clusters to detect two communities in the graph with most dissimilar
properties, and in another case, we have detected more than 2 clusters to analyze the
properties of more than 2 clusters. This experiment is done to validate the potential of
trust mechanism to detect fake clusters.
After the desired number of clusters are detected, their properties are analyzed based
on trust mechanism to label them as clusters of real users or clusters offake users. Figure 1
shows the detected clusters and Table 1 lists the properties of these detected clusters.
(a) Cluster 1 (b) Cluster 2 (c) Cluster 3
Fig. 1. Three clusters detected from social network graph which have the most similar
properties within the cluster and most different properties with another cluster.
Table 1. Properties of the detected clusters

Properties Cluster 1 Cluster 2 Cluster 3
Trust 0.546827 0.566929 0.798305
Density 0.019875 0.006766 0.621052
Results of the clustering and analysis of the properties show that the density of
cluster 3 is very high as compared to the density of cluster 1 and cluster 2. Also, the
average trust value of cluster 3 is also high. This shows that the nodes in cluster 3 have
high trust connections within the densely connected small community and have
decidedly fewer connections to outside communities. So, based on trust, the probability
of users in cluster 3 being fake is high as compared to cluster 1 and two because
gaining high trust only from a small community in less time is very rare.
To further validate our point, we have used an evaluation metric based on density,
time and trust to categorize the detected clusters. Cluster 1 is sparsely connected
average trust cluster of users, present on the network for more than a year. Cluster 2 is
also a sparsely connected average trust cluster created at the instant of time. Cluster 3 is
densely created high trust cluster created at the instant of time. All these detected
clusters can either be clusters of real users or clusters of fake users.
To show how variation of trust can able to detect fake clusters accurately, we have
plotted six months trust data of the clusters.
1
0.9
0.8
0.7
0.6
Trust
0.5
0.4
0.3
0.2
0.1
0
1 2 3 4 5 6
Months
BadCluster GoodCluster(Old) GoodCluster(New)
Fig. 2. Change of trust values over time for different clusters. Clusters with high trust value
during creation and fewer fluctuations over time are more probable of being clusters of fake users.
Above Fig. 2 shows the graph of trust values of the clusters over time. The graph
validates our point of fake users being gaining very high trust in the creation of the
cluster and maintaining it over time while real users tend to gain trust more gradually
and have more fluctuations.
5 Conclusion
We have developed a trust-based framework to detect clusters of fake users on social

networks. We have used trust-based metrics to capture the local and human knowledge
about the connections along with a machine learning algorithm. Based on the trust of
clusters and variation of trust over time we have analyzed the clusters of the social
network. Experiments are performed on the randomly generated social network, and
results validate the potential of trust-based security mechanism to detect fake user
accounts.
650 D. Kaur et al.
Acknowledgments. This work was partially supported by the National Science Foundation
under Grant No. 1547411 and by the U.S. Department of Agriculture (USDA) National Institute
of Food and Agriculture (NIFA) (Award Number 2017-67003-26057) via an interagency part-
nership between USDA-NIFA and the National Science Foundation (NSF) on the research
program Innovations at the Nexus of Food, Energy and Water Systems.
References
1. Zhang, P., Durresi, A.: Trust management framework for social networks. In: 2012 IEEE
International Conference on Communications (ICC), Ottawa, ON, pp. 1042–1047 (2012)
2. Zhang, P., Durresi, A., Barolli, L.: Survey of trust management on various networks. In:
2011 International Conference on Complex, Intelligent, and Software Intensive Systems,
Seoul, pp. 219–226 (2011)
3. Ruan, Y., Durresi, A., Uslu, S.: Trust assessment for internet of things in multi-access edge
computing. In: 2018 IEEE 32nd International Conference on Advanced Information
Networking and Applications (AINA), Krakow, pp. 1155–1161 (2018)
4. Ruan, Y., Durresi, A., Alfantoukh, L.: Using Twitter trust network for stock market analysis.
Knowl.-Based Syst. 145, 207–218 (2018)
5. Ruan, Y., Zhang, P., Alfantoukh, L., Durresi, A.: Measurement theory-based trust
management framework for online social communities. ACM Trans. Internet Technol. 17
(2), 24 (2017). Article 16
6. Ruan, Y., Durresi, A.: A survey of trust management systems for online social communities
– trust modeling, trust inference and attacks. Knowl.-Based Syst. 106, 150–163 (2016)
7. Ruan, Y., Durresi, A., Alfantoukh, L.: Trust management framework for internet of things.
In: 2016 IEEE 30th International Conference on Advanced Information Networking and
Applications (AINA), Crans-Montana, pp. 1013–1019 (2016)
8. Cao, Q., Sirivianos, M., Yang, X., Pregueiro, T.: Aiding the detection of fake accounts in
large scale social online services. In: Proceedings of the 9th USENIX Conference on
Networked Systems Design and Implementation (NSDI 2012), p. 15. USENIX Association,
Berkeley (2012)
9. McAuley, J., Leskovec, J.: Learning to discover social circles in ego networks. In: Pereira,
F., Burges, C.J.C., Bottou, L., Weinberger, K.Q. (eds.) Proceedings of the 25th International
Conference on Neural Information Processing Systems - Volume 1, (NIPS 2012), vol. 1,
pp. 539–547. Curran Associates Inc. (2012)
10. Fire, M., Katz, G., Elovici, Y.: Strangers intrusion detection-detecting spammers and fake
profiles in social networks based on topology anomalies. Hum. J. 1(1), 26–39 (2012)
11. Ferrara, E., Varol, O., Davis, C., Menczer, F., Flammini, A.: The rise of social bots.
Commun. ACM 59(7), 96–104 (2016)
12. Xiao, C., Freeman, D.M., Hwa, T.: Detecting clusters of fake accounts in online social
networks. In: Proceedings of the 8th ACM Workshop on Artificial Intelligence and Security.
ACM (2015)
13. Zhang, Z., Kulkarni, S.R.: Detection of shilling attacks in recommender systems via spectral
clustering. In: 2014 17th International Conference on Information Fusion (FUSION). IEEE
(2014)
14. Kagan, D., Elovichi, Y., Fire, M.: Generic anomalous vertices detection utilizing a link
prediction algorithm. Soc. Netw. Anal. Min. 8(1), 27 (2018)
15. Hamlili, A.: Intelligibility of Erdös-Rényi random graphs and time varying social network
modeling. In: ICSDE (2017)
Machine Learning Based Approach
to Detect Wormhole Attack in VANETs
Pranav Kumar Singh1,2 , Rahul Raj Gupta1 , Sunit Kumar Nandi1,3 ,

and Sukumar Nandi1(B)
1
Department of CSE, Indian Institute of Technology Guwahati,
Guwahati 781039, India
snghpranav@gmail.com, rahulrg.raj@gmail.com, sunitnandi834@gmail.com,
sukumar@iitg.ac.in
2
Department of CSE, Central Institute of Technology Kokrajhar,
Kokrajhar 783370, Assam, India
3
Department of CSE, National Institute of Technology, Arunachal Pradesh,
Yupia 791112, India
Abstract. VANET is one of the key enabling technologies for connected

and autonomous vehicles (CAVs). In a vehicular plane of the VANETs,
vehicles communicate with each other for various safety and non-safety
applications. In future, the autonomous CAVs communication will not
only be of type broadcast but also unicast in nature using multi-hop
communication. Both broadcast and unicast multi-hop communication
of VANETs are vulnerable to various types of attacks such as Denial
of Service (DoS), message falsification, Sybil attack, Greyhole, Black-
hole, and Wormhole attack. This paper aims to utilize the power of
machine learning to detect wormhole attack in multi-hop communica-
tion of VANETs. Although various mechanisms have been proposed in
the literature to detect this attack, the ML-based approach for wormhole
has not been explored. To model the attack in VANET, we create a sce-
nario of multi-hop communication using AODV routing protocol on NS3
simulator that uses the mobility traces generated by the SUMO traf-
fic simulator. We run the simulation and collect the statistics generated
using the flow monitor tool. These collected traces are preprocessed, and
then k-NN and SVM are applied on this preprocessed file to make the
model learn of wormhole attack. The performance of these two machine
learning models is compared in terms of detection accuracy and four
alarm types. Our study demonstrates that ML is a powerful tool, which
can help deal with such attacks in a multi-hop communication of future
generation CAVs.
Keywords: Wormhole · SVM · k-NN · SUMO · VANETs
1 Introduction
Vehicular ad-hoc networks (VANETs) is a subclass of mobile ad-hoc net-
works (MANET), with characteristics of high mobility, frequently changing net-
work topology and additional complexities [5]. VANETs facilitate many safety
https://doi.org/10.1007/978-3-030-15035-8_63
652 P. K. Singh et al.
and non-safety applications to connected and autonomous vehicles (CAVs) by

enabling them to communicate with other vehicles and wireless entities around
them. Such communications are termed vehicle-to-everything communications
(V2X), which provide road safety, reduces fuel consumption and carbon emis-
sion, and improves driving comforts. However, benefits never come free, there
will always be challenges. The biggest challenge of VANETs is dealing with a
wide range of security attacks. In various studies, it has been highlighted how
different layers of the VANET protocol stack is vulnerable to security attacks
from insider and outsider adversaries.
A wormhole attack [12] is one such security attack at the network layer of
VANETs which can severely affect the unicast multihop communication based on
popular routing protocols such as Ad Hoc On-Demand Distance Vector (AODV)
[18] and Dynamic Source Routing (DSR) [14]. This attack typically requires the
presence of at least two malicious vehicles in the network. These two malicious
vehicles need to be geographically separated and connected via a tunnel. Such
malicious insiders intend to deceive legitimate vehicles and purport them as if
they are good neighbors and can help in finding optimized route.
While the existing security framework based on public key infrastructure
(PKI) is immune from outside attacks, the detection of malicious behavior by
insiders is still in its infancy. In this paper, we present a machine learning based
approach to detect wormhole attack in VANETs. The work in this paper is based
on the wormhole attack on AODV routing protocol. We created the scenario on
simulation platform. We first train the machine learning models and then use
it to predict the behavior of each vehicle in the vehicular plane based on the
rules that models have learned. We compare the performance of two ML models
Support Vector Machine (SVM) and k-Nearest Neighbor (k-NN) to demonstrate
their effectiveness in this context.
2 Related Work
In 2003, the Hu et al. [11] introduced the wormhole attack and discussed how
severe this attack is in mobile adhoc networks. It is challenging to defend against
and can form a serious threat especially against many adhoc routing protocols.
In [4,20], the authors provided the details of security threats to adhoc routing
protocols and highlighted that the wormhole attack is one of the most powerful
attacks. As a result, various techniques have been proposed to detect, avoid and
defend the wormhole attacks. The detection techniques are categorized into three
categories [1]: hardware-based, anomaly-based, and guard node-based mecha-
nisms.
There are some more methods exist in the literature to detect wormhole
attacks in VANETs. However, most of these proposals either have issues in terms
of communication overhead and delay in detection, or they require time synchro-
nization, localization, and special hardware to achieve the goal. We believe that
ML can play an important role here to address the issues of the existing system
and can help detect with better accuracy without incurring additional overhead
and complexity.
Machine Learning Based Approach to Detect Wormhole Attack in VANETs 653
ML-based mechanism is not new to the attack detection in VANETS. There

are few good studies [8], [15], [22], [17] exist in the literature, which applied ML
to detect attacks, misbehavior and intrusions at different planes of the VANETs
such as In-Vehicle, Vehicular plane and Infrastructure plane.
Kang et al. [15] proposed a deep neural network (DNN) based mechanism
to detect attack on the in-vehicle system. They deployed intrusion detection
system (IDS) to detect attacks in a controller area network (CAN). Taylor et al.
[22] proposed Long Short-Term Memory (LSTM) neural network based anomaly
detector which detects attacks on CAN bus. Loukas et al. [17] proposed deep
learning based IDS which can be deployed at cloud to detect cyber-physical
attacks inside the vehicle. These works applied ML in detecting attacks at in-
vehicle systems.
In [8], the authors proposed ML-based approaches to classify the node’s
behavior, i.e. whether the communicating nodes in a vehicular plane is honest
or malicious. They implemented different types of misbehaviors and used Naive
Bayes, IBK, J-48, Random Forest and Ada Boost1 classifiers to classify the
behavior. Ali et al. [2] proposed an ML-based IDS to detect grey hole and rush-
ing attacks in a vehicular plane. They tested Support Vector Machines (SVM)
and Feed Forward Neural Networks (FFNN) for attack detection.
In [19], the authors proposed a ML-based solution to detect misbehavior at
the infrastructure plane of Cooperative-ITS. They used the predictive capabil-
ities of Multi-Layer Perceptron (MLP) and LSTM models to detect the false
information propagated by compromised traffic detection units.
Various ML-based mechanisms have been used to detect attacks at different
plane. However, ML-based detection of wormhole attack in a vehicular plane is
not available and we believe that machine learning has great potential to detect
this attack. To this end, our work is one such contribution.
3 Models: System, Communication and Adversary

In this section, we discuss preliminaries of our work that includes system model,
communication, adversary, attack and threat models.
3.1 System Model

The Fig. 1 depicts the system model of a VANET which consists of three planes:
Vehicular Plane, roadside unit (RSU) plane and Services plane.
In the vehicular plane of the VANET, vehicles communicate with each other
using vehicle-to-vehicle communication (V2V). The OBUs of vehicles have the
communication protocol stacks of the corresponding regions for example IEEE
WAVE-based stack in the USA. The RSU plane of the VANET facilitates vehicle-
to-infrastructure (V2I), infrastructure-to-infrastructure (I2I) and infrastructure-
to-vehicle (I2V) communications. There are various types of services, which can
be deployed at this plane of VANET such as ITS-based, payment, infotainment,
Internet, cloud-based, etc. It also constitutes the trusted authority (TA) or cer-
tificate authority (CA) of the PKI-based authentication system of VANETs.
Fig. 1. VANETs system model
3.2 Communication Model
In this work, we consider the Dedicated Short Range Communication (DSRC)

as radio access technology for V2V communication. The DSRC is specifically
designed for VANETs and standardized as IEEE 802.11p [13]. The protocol
stacks of VANETs support different communication strategies as shown in
Fig. 2a. These can be of mainly three types: unicast, multicast or geocast and
broadcast communication. In this work, we emphasize on unicast communica-
tion in which data communication from a source node to a destination node is
performed via multi-hop wireless communication if they are not in their direct
communication range. Two very popular types of routing strategies adopted for
data forwarding in multi-hop scenarios are: topology-based, and position-based
routing. We consider AODV protocol, which is a reactive routing protocol and
falls under the topology-based category.
(a) Communication Scenarios in (b) AODV Routing in

VANETs VANETs
Fig. 2. Communication model
Figure 2b shows the flow of different packets of AODV routing protocols.

Basic steps involved in AODV are:
– Each node keeps track of its neighbors by listening hello messages that each
node broadcast at defined interval.
– If a node wants to send data to the node which is not its neighbor, then it
broadcast RREQ (Route Request) message.
– Every neighbor which receives RREQ, either return RREP (Route Reply) (if
know the route to destination or destination itself) or rebroadcast RREQ.
– When destination is found, then RREP is sent via the reverse path.
– A higher sequence number signifies a fresh route, and the forward path is
used to route data packets.
– The RERR (Route Error) message allows AODV to adjust routes when the
node moves around.
3.3 Adversary, Attack and Threat Model
The wormhole attack can be launched by a malicious node, which may be either
an outsider node or an insider node. The outsider node does not have crypto-
graphic materials such as certificate and keys are given by the CA whereas the
insider possesses the required certificates and keys. In this work, we consider
internal adversaries who in real life may be created by compromising legitimate
nodes. These compromised nodes can have Byzantine behavior and can also
cooperate or collude amongst themselves.
The adversaries launches wormhole attacks by exploiting the vulnerabilities
of popular ad-hoc routing protocols, such as DSR and AODV. The common
methods of launching wormhole attack are [16]: packet encapsulation, out-of-
band channel (OBC), high power transmission, packet relay and protocol devi-
ations. In this work, we consider wormhole using packet encapsulation. Due to
space constraint we are not providing the details of this attack. However, authors
can refer works [1,16] for details of all these attacks.
The wormhole attack has been considered as one of the severe security
attacks, which can severely affect the multihop communication in VANETs. The
malicious nodes who collude to from the tunnel, attract routes through them and
use it to pass the routing traffic. Variety of attacks can be launched on the data
traffic flowing through the wormhole tunnel, such as dropping the data packets
selectively. This attack can also prevent nodes to discover legitimate routes and
thus disrupt normal functionality of VANETs. Moreover, it can also affect data
aggregation, location-based wireless security, and clustering protocols [16].
4 Methodology
In our work, we propose a ML-based wormhole attack detection system that
determines the behavior of vehicles in VANETs. Our system uses the trace files
produced by simulator that consists both normal and abnormal behavior (under
wormhole attack) of nodes in VANETs. The steps below provide the details of
methodology adopted for our proposed system:
4.1 Simulation
We used two simulators to model the VANET scenario discussed in the previous
section: Simulation of Urban Mobility Model (SUMO) [6] as a traffic simula-
tor and NS3 as a network simulator [10]. We used the Manhattan mobility by
importing a map from OpenStreetMap (OSM) and converting to usable NS3 file
by generating trace using NETCONVERT, POLYCONVERT, randomTrips, and
traceExporter features of the SUMO. Figure 3a depicts the steps to produce the
SUMO configuration file from an OSM file [21]. Figure 3b shows the OSM file
and corresponding SUMO net file used in work. The selected road segment is of
size 3 km2 and each vehicle is of length 5 meters and max speed of 15 m/s.
The mobility trace file generated by the SUMO is fed into the NS3 simulator
using its ns2mobilityhelper module. The Table 1 shows the parameters that we
used for our simulation scenario.
(a) Steps to Generate Mobility Trace

using SUMO (b) OSM to SUMO
Fig. 3. Traffic simulation
4.2 Dataset Generation

From the simulation discussed in the previous step, we generate a dataset to
be used for learning the attack. In order to prepare the dataset, we dump all
the data link statistics into a file using flow monitor tool of NS3. The different
features that we extract are Source and destination IP, Transmitted and received
Bytes, Bytes Dropped, Delay Sum, Jitter Sum, Lost Packets, Packets Dropped,
F irstRxBytesT ime − F irstT xBytesT ime, and Throughput.
To generate a wormhole attack scenario in our simulation, we modified AODV
routing protocol module in NS3. We declared two nodes as attacking nodes and
modified AODV module treats these two nodes as wormhole attacking nodes,
which creates a high-speed tunnel between them. Thus, whenever any of these
nodes receive any RREQ packet, it uses its tunnel and replays the packet to
other attacking nodes which further delivers the request to the destination node.
Table 1. Simulator parameter and values
Parameter Value
Simulator NS 3.26
Simulation time 400 s
Number of vehicles 40
Traffic type Constant Bit Rate (CBR)
Topology 3000 m ∗ 3000 m
Transport protocol UDP
Packet size 512 Bytes
Routing protocol AODV
Channel type SCH
Number of road lanes 2
Radio propagation model Two ray ground
MAC protocol IEEE 802.11p
Max speed 15 m/s
Mobility model Manhattan mobility model
Hence, claims itself as an intermediate node in the path from source to destina-
tion node. For generating the dataset, we simulate the network multiple times
with a different pair of nodes as attacking nodes. We label the transmission
occurring between the two attacking nodes as abnormal and the rest as normal.
Table 2. Dataset characteristics
Parameter Value
Number of features 11
Classes 2 (0-Normal, 1-Abnormal)
Number of instances 11444
Class label 0 10000
Class label 1 1444
Training size 8010
Testing size 3434
Missing values NA
4.3 Preprocessing and Feature Selection
After extracting the features from flow monitor in the previous stage, we need to
assure that each class has balanced class distribution. In our prepared dataset, we
have more than 98% instances labeled normal. We use downsampling technique
to overcome this issue which randomly eliminates majority class instances from
the dataset. Table 2 shows the characteristics of the final dataset to be used for
learning the attack.
From the feature set, some features may not be correlated with the class
labels. The feature selection algorithm [9] selects a subset of features which gives
maximum accuracy. This method helps in finding the best subset of features
which directly affects the behavior of the system. The straightforward approach
in finding the best subset of features is to apply learning model for every possible
subset of features. The subset on which the model gives the maximum accuracy
is the required set of features. The straightforward approach has a very large
search space. There are many other standard feature selection algorithm. In this
work, we use the Sequential Forward Selection (SFS) algorithm and Sequential
Backward Selection (SBS) algorithm. These two algorithms are greedy in nature.
Figure 4a and 4b show the variation of accuracy with a number of features.
From the experiment, the feature selection algorithm returns the set of fea-
tures, which gives maximum accuracy are Source and destination IP Address,
Transmitted and Received Bytes, Dropped Bytes, F irstRxBytesT ime −
F irstT xBytesT ime and Throughput.
Fig. 4. Feature selection
4.4 Machine Learning Models for Detection

In this stage, we apply the Support Vector Machines (SVM) and k-nearest neigh-
bors (k-NN) machine learning models on the self created dataset with the fea-
tures selected in previous stage.
SVM is a supervised learning algorithm of machine learning models that
solves both classification or regression challenges [7]. It is one of the discrimi-
native classifiers that tries to differentiate data points into classes using hyper-
planes. The k-NN algorithm [3] can also solve both classification and regression
problems, but mainly used for classification tasks. For each data point, we con-
sider k nearest training data points and predict the most occurring class for a
test data point. k is a hyperparameter which represents the number of training
points to be considered for classifying test data point.
5 Results
In this section, we discuss the results that we achieved after the experiment.
We split the dataset into training and testing data. The training data consists
of 70% of total instances, and testing data contains 30% of total instances. For
KNN, the value of k, the number of nearest neighbors is 4. The value of k is set
to 4 based on accuracy that we got in experiment. The accuracy of the model is
the number of correctly classified test instances out of total test instances. Both
the SVM and k-NN performed well and detection the attack with accuracy of
99%.
We compare the performance of the learning models used to learn the attack
on the basis of accuracy and the different alarm rates. The four different alarm
types are True Positive (TP), True Negative (TN), False Positive (FP) and False
Negative (FN). The different alarm rates for wormhole attack by two models are
shown in Table 3. From learning models, we expect a high accuracy and a low
false alarm rate. From the results, we observe that both the models perform well
for detecting wormhole attack with the high accuracy and a low false alarm rate.
Table 3. Comparison of SVM and KNN results
Alarm type SVM KNN

TP 0.9906 0.9953
TN 0.9986 0.9913
FP 0.0014 0.0087
FN 0.0094 0.0047
6 Conclusion
In this paper, a machine learning model-based approach is proposed to detect
wormhole attack in VANETs. We demonstrated with our methodology that how
effective ML is in detecting wormhole attack in the network. To get the best set
of features, we use feature selection algorithm. The method proposed to use a
self-created dataset to learn the attacks. The described model detects the attack
with high accuracy and low false alarm rate. ML has great potential and can be
utilized in detecting attacks at different layers of the VANET protocol stack.
References
1. Albouq, S.S., Fredericks, E.M.: Detection and avoidance of wormhole attacks in
connected vehicles. In: Proceedings of the 6th ACM Symposium on Development
and Analysis of Intelligent Vehicular Networks and Applications, pp. 107–116.
ACM (2017)
2. Ali Alheeti, K.M., Gruebler, A., McDonald-Maier, K.: Intelligent intrusion detec-
tion of grey hole and rushing attacks in self-driving vehicular networks. Computers
5(3), 16 (2016)
3. Altman, N.S.: An introduction to kernel and nearest-neighbor nonparametric
regression. Am. Stat. 46(3), 175–185 (1992)
4. Argyroudis, P.G., O’mahony, D.: Secure routing for mobile ad hoc networks. IEEE
Commun. Surv. Tutor. 7(1–4), 2–21 (2005)
5. Bakhouya, M., Gaber, J., Lorenz, P.: An adaptive approach for information dis-
semination in vehicular ad hoc networks. J. Netw. Comput. Appl. 34(6), 1971–1978
(2011)
6. Behrisch, M., Bieker, L., Erdmann, J., Krajzewicz, D.: SUMO–simulation of urban
mobility. In: The Third International Conference on Advances in System Simula-
tion (SIMUL 2011), Barcelona, Spain, vol. 42 (2011)
7. Cortes, C., Vapnik, V.: Support-vector networks. Mach. Learn. 20(3), 273–297
(1995)
8. Grover, J., Prajapati, N.K., Laxmi, V., Gaur, M.S.: Machine learning approach
for multiple misbehavior detection in VANET. In: International Conference on
Advances in Computing and Communications, pp. 644–653. Springer (2011)
9. Guyon, I., Elisseeff, A.: An introduction to variable and feature selection. J. Mach.
Learn. Res. 3(Mar), 1157–1182 (2003)
10. Henderson, T.R., Lacage, M., Riley, G.F., Dowell, C., Kopena, J.: Network simu-
lations with the ns-3 simulator. SIGCOMM Demonstr. 14(14), 527 (2008)
11. Hu, Y.C., Perrig, A., Johnson, D.B.: Packet leashes: a defense against wormhole
attacks in wireless networks. In: Twenty-Second Annual Joint Conference of the
IEEE Computer and Communications, INFOCOM 2003. IEEE Societies, vol. 3,
pp. 1976–1986 IEEE (2003)
12. Hu, Y.C., Perrig, A., Johnson, D.B.: Wormhole attacks in wireless networks. IEEE
J. Sel. Areas Commun. 24(2), 370–380 (2006)
13. IEEE Std.: IEEE standard for information technology– local and metropolitan area
networks – specific requirements– Part 11: wireless LAN medium access control
(MAC) and physical layer (PHY) specifications amendment 6: wireless access in
vehicular environments, pp. 1–51, July 2010
14. Johnson, D.B., Maltz, D.A., Broch, J., et al.: DSR: the dynamic source routing
protocol for multi-hop wireless ad hoc networks. Ad Hoc Netw. 5, 139–172 (2001)
15. Kang, M.J., Kang, J.W.: Intrusion detection system using deep neural network for
in-vehicle network security. PloS One 11(6), e0155781 (2016)
16. Khalil, I., Bagchi, S., Shroff, N.B.: Liteworp: detection and isolation of the worm-
hole attack in static multihop wireless networks. Comput. Netw. 51(13), 3750–3772
(2007)
17. Loukas, G., Vuong, T., Heartfield, R., Sakellari, G., Yoon, Y., Gan, D.: Cloud-
based cyber-physical intrusion detection for vehicles using Deep Learning. IEEE
Access 6, 3491–3508 (2018)
18. Perkins, C., Belding-Royer, E., Das, S.: Ad hoc on-demand distance vector (AODV)
routing. Technical report (2003)
19. Singh, P.K., Dash, M.K., Mittal, P., Nandi, S.K., Nandi, S.: Misbehavior detec-
tion in C-ITS using deep learning approach. In: 18th International Conference on
Intelligent Systems Design and Applications (ISDA). Springer (2018)
20. Sanzgiri, K., Dahill, B., Levine, B.N., Shields, C., Belding-Royer, E.M.: A secure
routing protocol for ad hoc networks. In: 10th IEEE International Conference on
Network Protocols, pp. 78–87. IEEE (2002)
21. Singh, P.K., Sharma, S., Nandi, S.K., Nandi, S.: Multipath TCP for V2I commu-
nication in SDN controlled small cell deployment of smart city. Veh. Commun. 15,
1–15 (2018)
22. Taylor, A., Leblanc, S., Japkowicz, N.: Anomaly detection in automobile control
network data with long short-term memory networks. In: 2016 IEEE International
Conference on Data Science and Advanced Analytics (DSAA), pp. 130–139. IEEE
(2016)
Secure Peer-to-Peer Communication
Based on Blockchain
Kahina Khacef(B) and Guy Pujolle(B)
Sorbonne University, 4 Place Jussieu, 75005 Paris, France

{kahina.khacef,guy.pujolle}@lip6.fr
Abstract. Nowadays, electronic messaging is the most used network

application, and the authentication between users is a vital property.
The most commonly employed approaches to ensure this property are
PKI and S/MIME email encryption protocols, but indeed they are fac-
ing multiples security threats, such as the MITM attack and EFAIL
attack. The blockchain is an innovative technology that overcomes these
threats and allows to decentralize sensitive operations while preserving
a high level of security. It eliminates the need for trusted intermediaries.
The blockchain is accessible to all network nodes and keeps track of all
transactions already made. The goal of our work is to propose a secure
messaging solution based on the blockchain technology. In this paper, we
explain why blockchain would make communications more secure, and
we propose a model design for blockchain-based messaging maintaining
the performance and security of data recorded on the blockchain, using
a smart contract to verify the identities and their associated public keys,
and validate the user’s certificate. The system is entirely decentralized
and allows users to exchange messages securely.
1 Introduction
Public Key Infrastructure (PKI) is a significant component to resolve authen-
tication in networks and provides guarantees to trust a certificate signed by a
certification authority (CA). A new concept named web of trust for his Pretty
Good Privacy (PGP) encryption program uses PKI to provide confidentiality
with encryption, authentication via the signature and web of trust via identity
validation from peers. The certificates authenticate the public keys and allow
you to perform cryptographic operations, such as encryption and digital signing.
As authentication and identity validation is centralized in the PKI, that creates
single points of failure.
The blockchain technology was designed to make the transaction more reliable.
Blockchain technology is a potential solution to achieve data integrity, relying on
cryptography to provide tamper-resistance. Blockchain can be utilized in such
cases to achieve the secure communication and integrity of data. The blockchain is
decentralized, and no centralized authority can approve transactions. All network
participants must reach a consensus to validate transactions in a secure way, and
previous records cannot be changed. A very high cost must be spent if someone
https://doi.org/10.1007/978-3-030-15035-8_64
Secure Peer-to-Peer Communication Based on Blockchain 663
wants to alter previous records. External attackers would have to gain access to
every computer in the network that hosts the blockchain database at the same
time to manipulate it, which is as practically impossible [1].
The cryptocurrency Bitcoin was created by an unknown person using the
alias Satoshi Nakamoto in 2008. Bitcoin generates and broadcast transactions
to the Blockchain. Once validated, the transactions are propagated through the
network and added into a block. Once a block is full, the block is appended
to the Blockchain by performing a mining process. Miners try to solve a hard
cryptographic puzzle named Proof of Work (PoW), and the node that solves the
puzzle first adds the new block to the Blockchain. Bitcoin is based on decen-
tralized trust. The trust is achieved as an emergent property from the interac-
tions of different participants in the bitcoin system [2]. The data stored on the
blockchain cannot be hacked, modified or deleted. The immutability of data in
this blockchain is strongest when the chain is long [3].
1.1 Motivations
Blockchain has been of great interest to engineers and investors because of its
immense commercial potential and its use in applications as diverse as a cryp-
tocurrency.
Several layering schemes of new features on bitcoin are proposed. A proposal
to create a decentralized domain name service over Bitcoin appeared as BitDNS,
eventually evolved into Namecoin, the first altcoin [4]. Bitcoin has the most sub-
stantial amount of computational power securing the blockchain data. However,
it’s hard to introduce new functionality to bitcoin because that requires consen-
sus breaking changes. Transactions in bitcoin did not include a standard method
for carrying a data payload, and the blockchain grows at an exponential rate [5]
that extends pressures upward on storage space and importantly network band-
width. Due to the security concerns, only a handful of functions are permitted
in standard transactions [6].
Peer to peer (P2P) systems help overcome many problems that go beyond
traditional client-server approaches, but these features also introduce new issues
such as how to establish the trust relationship within P2P networks [7].
The Ethereum blockchain is a platform for decentralized applications called
smart contracts. Because Ethereum addresses are unique identifiers whose own-
ership does not change, their activity can be tracked and analyzed [8]. A smart
contract is an executable code that runs on top of the blockchain, allows run-
ning an agreement between two parties automatically, without one of the parties
being able to obstruct its execution [9].
At the time of the digital economy, data is brought to transit more and
more between companies, from a client to a supplier, moving from one cloud to
another with the advent of virtualization and containers. Our work describes the
potential for applying blockchain to assure the traceability of data, certificate
individual and secure messaging based on the blockchain technology.
664 K. Khacef and G. Pujolle
1.2 Contribution and Organisation of Paper
In this contribution, we are interested in the security of the blockchain for mes-
saging, which can be an email, a website or some other form of message. We
propose a protocol that performs encrypted messaging on the public blockchain.
The goal of our contribution is to achieve the three functionalities achieved by
PKI using the blockchain as the database to store public keys, digital signature,
and peer information, allowing each entity of the network to validate information
about every other node in the network.
The main contribution of this paper is: Sect. 2 a presentation of the benefits
of blockchain in cybersecurity, Sect. 3 briefly description of the current state of
the art for Blockchain based PKI, and in Sect. 4 we present our proposed solution
with a smart contract to use the legitimate secure certificate.
2 Blockchain
A blockchain is essentially a fault-tolerant distributed database of records of

a public ledger of all transactions that have been executed and shared among
participating parties [10]. Each transaction in the public ledger is verified by a
consensus of a majority of participants in the system. Blockchain provides a tool
for increasing data integrity using a combination of cryptography and consensus
[11]. As such, there is no central server point of control. The information is
stored in blocks, Each block contains transactions, a timestamp and a link to
the previous block, and is cryptographically protected. What makes it secure is:
• Public key infrastructure: It uses public/private key encryption and data

hashing to store and exchange data safely.
• Distributed ledgers: There is no central authority to hold and store the
data, it removes the single point of failure.
• Peer to peer Network: The communication is based on the P2P network
architecture and inherits the decentralized characteristics.
• Cryptography: Blockchain uses a variety of cryptographic techniques, hash
functions, Merkle trees, public and private key [12]. It is difficult to alter the
blockchain, to make a modification, it is necessary to succeed in a simultane-
ous attack on more than 51% of the participants.
• Consensus Algorithm: The rules which the nodes in the network follow to
verify the distributed ledger. Consensus algorithms are designed to achieve
reliability in a system involving multiple unreliable nodes. A consensus of the
nodes validates the transactions. Choice of consensus algorithm has signifi-
cant implications for performance, scalability, latency and other Blockchain
parameters. The consensus algorithms must be fault-tolerant [13].
• Transparency: Each transaction is visible to anyone with access to the sys-
tem. If an entry can not be verified, it is automatically rejected. The data is
therefore wholly transparent. Each node of a blockchain has a unique address
that identifies it. A user may choose to provide proof of identity to others.
2.1 How Public Blockchain Solves PKI Problems?

In blockchain, the shared data is the history of every transaction ever made. The
ledger is stored in multiple copies on a network of computers, called “nodes.”
Each time someone submits a transaction to the ledger, the nodes check to make
sure the transaction is valid. The transaction is being created in the chained data
structure of blockchain, a new timestamp will be recorded at the same time, and
any modification of data created before will not be allowed anymore.
A block is a data structure which consists of a header and a list of trans-
actions. Each transaction is generally formed as trx := (M, Signature) where
M := (PKSender, receiver, data); PKSender denotes the public key of a sender
(address of the sender is derived from the PKSender), and receiver is the address
of a receiver, Signature := Sign SKSender (H(M)) where SKSender is the private
key of the sender and H() is a secure hash function. A subset of them competes
to package valid transactions into “blocks” and add them to a chain. The owners
of these nodes are called miners and the miners who successfully add new blocks
to the chain earn a reward.
A cryptographic hash unique to each block and a consensus protocol makes
it tamper proof [13]. There are different types of the blockchain, and there are
several consensus mechanisms used in the blockchain, we present the two most
used consensus proof of work (PoW) and proof of stake (PoS).
2.1.1 Types of Blockchains

1. Public blockchain: A blockchain that anyone in the world can read, can
send transactions to and expect to see them included if they are valid. This
means anyone can become part of the network and participate in the consen-
sus process making them permissionless. There is no way to censor transac-
tions on the network nor change transactions retrospectively [14]. The content
of the blockchain can be trusted to be correct. Public blockchains are, how-
ever, very inefficient. The more computing power is required to support trust.
So, an attacker would need to acquire 51% of the network’s computing power
to change an entry in the blockchain. (e.g., Bitcoin, Ethereum, ZCash).
2. Consortium blockchain: It is a blockchain where a pre-selected set of nodes
control the consensus process.
3. Private blockchain: A blockchain where access permissions are more lightly
controlled, where rights to modify or even read the blockchain state restricted
to a few users, where only known nodes are allowed to participate in the
network. Ideally, it is internal for an organization. The writes permissions are
kept centralized to one organization. Private blockchain reduces counterparty
risk by enabling the exchange of data without the intermediation of third
parties.
4. Permissioned Blockchain: It is a blockchain where we can allow specific
actions to be performed only by specific addresses. The participants in the
network can restrict who can participate in the consensus mechanism and
who can create a smart contract and give the authority for some participants
to provide the validation of blocks of transactions. A control access layer into
the blockchain nodes is used. However, raise their questions, Who has the
authority to grant permission? A permission blockchain may make its owners
feel more secure, giving the database rigorous security and privacy capabilities
but can see as violating the idea of blockchain because only some participants
have more control, which means they can make changes whether or not other
network participants agree.
2.1.2 Proof of Work

PoW is the consensus algorithm used in bitcoin. The protocol use of elliptic curve
cryptography for transactions signatures, and the elliptic curve digital signature
algorithm (EDCSA) [15] is used. The user needs to provide her public key and
her signature to prove ownership. A hash takes a lot of computing time and
energy to generate initially. It thus serves as a proof that the miner who added
the bloc to the blockchain did the computational work to earn. The hashes also
serve as the links in the blockchain, each block includes the previous block’s
unique hash.
If an attacker wants to tamper with the blockchain, he needs to control more
than 50% of the hashing power to become the first one to generate the new bloc.
2.1.3 Proof of Stake

There are no particularly heavy calculations in a PoS. PoS does not require
powerful and expensive hardware for maintenance of the blockchain and do not
consume a lot of electricity. PoS require significantly smaller resources than
PoW. The digital currency has the concept of coin age. Coin age of a coin is its
value multiplied by the period after it was created [16]. The probability that an
account will succeed in confirming the next block of transactions to add to the
blockchain is proportional to the amount of money that is on that account.
Hash(hash(Bprev), A, t) <= Sold(A) M/D, with Sold(A), is the balance of
address A, and t is the timestamp. Unlike the PoW, the only variable that the
user can change is the timestamp.
3 Related Work
In this section, we present the works made use Blockchain for secure transactions
between users. Since BitDNS [4], numerous approaches are proposed.
Namecoin it is the first system to build a decentralized naming system using
blockchain, the first altcoin from bitcoin with its blockchain [14]. Satoshi believed
that BitDNS should use its independent blockchain and event offered the first
proposal for merged mining as a way to secure blockchain. Namecoin was also
the first solution to zooko’s triangle producing a naming system that is simul-
taneously secure, decentralized, and human-meaningful. Namecoin suffers from
51% attacks [17], because of its insufficient computing power.
Blockstack uses the existing internet transport layer (TCP or UDP)
and underlying communication protocols and focuses on removing points of
centralization that exist at the application layer. The underlying blockchain of

Bitcoin limits the performance of Blockstack [18].
Certcoin [19] removes central authorities and uses the blockchain Namecoin as
a distributed ledger of domains and their associated public keys. Every certcoin
user stores the entires blockchain, and this causes two problems, the latency for
the controller and the security problems of merged mining used by Namecoin.
Emercoin [20]: Blockchain-based PKI that doesn’t remove central authorities
but uses Blockchains to store hashes of issued and revoked certificates. Emercoin
has the side benefit of optimizing network access by performing key and signature
verification on local copies of the blockchain.
However, all of these systems faced the same barrier. This consensus involves
significant energy, delay, and computation overhead because of high resource
demand for solving the PoW.
4 Contribution
4.1 The Proposed Model
The primary goal of our approach is to secure communications between enti-
ties of the network. The proposed model is to use the blockchain to validate
the user’s identity and to ensure trust between users for exchanging messages
with a high level of security. Each user must communicate only with the user’s
identity validated by the smart contract, and consider every other interaction
as malicious. Each user how want uses the system to communicate with others
must register their identity and their public key an store it on blockchain. To
achieve such a system, we can use Ethereum public blockchain that implements
smart contracts. The Ethereum network has up to 200,000 developers currently
working on the protocol and derived projects, ensuring that the protocol remains
Table 1. Notations
Notation Definition
Alice, Bob Entities
userA
Kpub Public key of Alice
userB
Kpub Public key of Bob
userA
Kpri Private key of Alice
userB
Kpri Private key of Bob
A
ID Id of Alice
IDB Id of Bob
userA
Spri Signature with private key of Alice
userA
Epub Encryption with public key of Alice
t Validity of the Public Key
T Time stamps
cutting edge, it has a level of security. The most frequently used notations in
our scheme are mentioned in Table 1.
Once the code conditions are validated in smart contract, it can run as sched-
uled. This execution is carried out in practice by sending a specific transaction
to the blockchain that causes the execution of the smart contract (Fig. 1).
4.1.1 Smart Contract

The smart contract is the code that is stored and executed on a blockchain.
Once a smart contract is deployed, users cannot modify it since the contract was
sented and validated. A user can send a transaction to the contact’s address,
and the transaction will be executed.
Each interaction with the smart contract is recorded as a “transaction” on
the blockchain. These transactions are grouped in a Merkle tree and stored in a
block on the blockchain.
4.1.2 Blockchain Registration

1. The user Alice generate a pair key using ECDSA algorithm, a public key
userA userA
Kpub and private key Kpri , then derives the identity IDA , as the hash
of the public key as shown in Fig. 2. Alice keeps the private key safely and
requests to register its identity with the corresponding public key into the
blockchain after to be verified and validate by the network.
Each public key have an associated timestamp.
2. Alice signs transactions with the corresponding private key and transfers
(IDA , Spri
userA userA
(Kpub , t, T )) to the blockchain.
3. The miner checks:
userA
a. Kpub (IDA , Spri
userA userA
(Kpub , t, T )).
A
b. That the ID has never previously been registered.
4. If verified, The certificate’s Alice (IDA , Spri

userA userA
(Kpub ), t, T ), is then stored
on the blockchain, with the following pieces of information: (Id Alice, Public
key of Alice, the validity of the public key and timestamp).
5. This is the registration process for all network entities.
Fig. 2. ID generation
4.1.3 Smart Contract-Based Verification

When user Bob wants to send a message to Alice, Bob only presents IDB previ-
ously recorded on the blockchain and IDA of Alice with a time-stamp T to the
blockchain. Each message must include the time.
1. Bob sents a transaction: TB = [IDB , IDA , T, SpriuserB

(IDB , IDA , T )].
The Smart contract receives the request from the user Bob:
2. The smart contract checks if the presented IDB and IDA exist on the
blockchain. The Smart contract reads and parses the two Alice and Bob
recordings.
a. It performs a Lookup of (Kpub , ID), If existed the output return true.
b. Once the validity of the public keys verified, the Smart contract checks
the validity of the timestamp and the signature of the transaction.
c. Finally, The smart contract validates the request and return true.
3. Bob sends a transaction TB1 = [IDB , IDA , T, Epub
userA
(IDB , IDA , T, Kpub
userB
)]
to Alice’s address.
4. Next, Alice checks the transaction with his private key and then sends to
Bob’s address TA2 = [IDA , IDB , T+1, Epub
userB
(IDA , IDB , T + 1, KpubuserA
)].
5. After receipt of the transaction by Bob, the same verification will be per-
formed, and mutual authentication is established between the two entities.
4.1.4 Send/Receive Message

Once mutual authentication is established, Alice uses Bob’s public key and Alice’s
private key to generate a shared secret. The shared secret can be generating using
the Elliptic-curve-Diffie-Hellman (ECDH) to encrypt Message. ECDH is a variant
of the Diffie-Hellman algorithm for elliptic curves. Once the shared key is gener-
ated, it is used to encrypt messages using a symmetric key algorithm [21].
4.2 Evaluation
4.2.1 Advantages
Our approach removes central authorities (CA) and uses the blockchain pub-
lic as a distributed ledger of identity and their associated public keys. We use
Blockchain to store public keys, digital signature, and peer information.
Once published, the smart contract code works precisely as programmed.
This is one of the main advantages of the platform, the code always interacts
as promised it cannot be falsified, and it never has a downtime. The system is
trust, transparent and traceable.
1. Confidentiality: Once the communication channel between users is secured,

peer to peer encryption between endpoints can be set and only authorized
users to have access to the messages exchanged.
2. Message integrity and Authentication: The blockchain checks the
validity of the signature, before being stored. Another person can not
change/modify the signed agreement or alter exchanged message during the
network transit. Each user has a certificate stored on the blockchain. The
smart contract checks the certificate and proves the identity of users. All
exchanged messages are signed with private keys associated with the public
key on certificates using the ECDSA algorithm.
3. Reliability: It is impossible to shut down all computers participating in the
blockchain simultaneously. As a result, this database is always online, and its
operation never stops.
4.2.2 Limitations
In blockchain, the smart contract is executed sequentially that affect the per-
formance of blockchain negatively. With the growing number of smart contract,
the blockchain will not be able to scale.
In practice, it is impossible to modify an existing contract in the blockchain
after it has been registered in it. The design phase of these contracts will, there-
fore, require special attention to avoid any future disappointment.
4.3 Implementation
The smart contract plays a vital role in executing and performing the agree-
ment among various users in the system. For implementing the proposed sys-
tem, a smart contract can be created by developing the codes that run on the
blockchain to execute the agreement signed by the nodes to validate the certifi-
cates of individuals. Smart contract promise low transaction fees compared to
traditional systems that require a trusted third party to execute an agreement.
The contract’s state is stored on the blockchain, and it is updated each time the
contract is invoked.
Before sending/receiving the message, the smart contract first checks the
identity and its associated public key, previously stored in the blockchain. Then,
it checks the validity of the timestamp. Finally, the smart contract verifies the
signature.
A smart contract can be developed with Solidity.
5 Conclusion
Our contribution benefits from an entirely decentralized architecture offering
inherent fault tolerance, redundancy, and transparency. We have firstly proposed
an approach that secure communication, it benefits of security properties of
blockchain public. It shows how to use to the immutability of the blockchain to
provide a solution to high problematics in the field of centralized PKI.
We plan to implement our proposal to base our results empirically and
demonstrate its viability.
Our next proposal is to set up an architecture with smart contracts to val-
idate, store and revoke the certificate on a public blockchain. The certificate
of the individual will contain, his address and public key, the address of smart
contract that issued it, and stored it in Off-chain.
References
1. Schuermann, F.: Bitcoin and beyond: a technical survey on decentralized digital
currencies. IEEE Commun. Surv. Tutor. 18, 2084–2092 (2016)
2. Bano, S., Sonnino, A., Bassam, M., Azouvi, S., McCorry, P., Meiklejohn, S.,
Danezis, G.: SoK: consensus in the age of blockchains, arxiv (2016)
3. Cachin, C., Vukoli, M.: Blockchain Consensus Protocols in the Wild (2016)
4. Loibi, A.: Namecoin, Seminar innovative Internettechnologien und mobilkommu-
nikation SS2014. IEEE (2014)
5. Hung, T., Svetinovic, D.: Data analysis of digital currency network: namecoin case
study. In: International Conference on Engineering of Complex Computer Systems
(2016)
6. Deepak, K., Shetty, S., Liang, X., Kamhoua, C., Njilla, L.: Consensus Protocols for
Blockchain-based Data Provenance: Challenges and Opportunities. IEEE (2017)
7. Mudliar, K., Parekh, H., Bhavathankar, P.: A comprehensive integration of national
identity with blockchain technology. In: 2018 International Conference on Com-
munication, Information and Computing Technology (ICCICT), 2–3 February,
Mumbai, India (2018)
8. Aung, Y.N., Tantidham, T.: Review of Ethereum: smart home case study. In: 2017
2nd International Conference on Information Technology (INCIT) (2017)
9. Ward, M.: Untangling Blockchain: A Data Processing View of Blockchain Systems,
Elsevier Information Security Technical report, pp. 89–92 (2006)
10. Tewari, H., Nuallain, E.: Netcoin: A Traceable P2P Electronic Cash System. IEEE
(2016)
11. Guy, Z., Oz, N., Alex, P.: Decentralized Privacy: Using Blockchain to Protect
Personal Data (2018)
12. Conti, M., Kumar, S., Lal, C., Ruj, S.: A Survey on Security and Privacy Issues of
Bitcoin. IEEE (2017)
13. Tosh, D.K., Shetty, S., Liang, X., Kamhoua, C., Njilla, L.: Consensus Protocols for
Blockchain based Data Provenance: Challenges and Opportunities. IEEE (2017)
14. Kalodner, H., Carlsten, M., Ellenbogen, P., Bonneau, J., Narayanan, A.: An empir-
ical study of Namecoin and lessons for decentralized namespace design. Blockchain
(2014)
15. Sankar, L.S., Sindhu, M., Sethumadhavan, M.: Survey of consensus protocols on
blockchain applications. In: International Conference on Advanced Computing and
Communication Systems (ICACCS 2017)
16. Mingxia, D., Xiaofeng, M., Zhe, Z., Xiangwei, W., Qijin, C.: A Review on Consen-
sus Algorithm of Blockchain. IEEE (2017)
17. Dai, F., Shi, Y., Meng, N., Wei, L., Ye, Z.: From bitcoin to cybersecurity: a com-
parative study of blockchain application and security issues/ICSAI, pp. 975–979
(2017)
18. Ali, M., Nelson, J., Shea, R., Freedman, M.J.: Blockstack: a global naming and
storage system secured by blockchain. In: USENIX Annual Technical Conference
(2016)
19. Fronknecht, C., Velicannu, G., Sophia, Y.: CertCoin: A Namecoin Based Decen-
tralized Authentication System (2014)
20. Khovayko, O., Shumilov, E.: EMCSSL Decentralized identity management, pass-
wordless logins, and client SSL certificates using Emercoin NVS (2016). http://
emercoin.com
21. Andrea, C.: ECDH and ECDSA (2015). https://andrea.corbellini.name/2015/05/
30/elliptic-curve-cryptography-ecdh-and-ecdsa
Developing the Analysis Tool
of Cyber-Attacks by Using CTI
and Attributes of Organization
Yusuke Kambara(B) , Yoshinori Katayama, Takanori Oikawa,

Kazuyoshi Furukawa, Satoru Torii, and Tetsuya Izu
Fujitsu Laboratories Ltd., 1-1, Kamikodanaka 4-chome,

Nakahara-ku, Kawasaki 211-8588, Japan
kambara-yusuke@fujitsu.com
Abstract. Cyber Threat Intelligence (CTI) is a source of useful infor-

mation for organizations to take countermeasures against cyber-attacks.
The process of using CTI is not automatic and requires human interven-
tions, because we have to (1) check the CTI data to avoid obstructing
business before executing countermeasures and (2) identify the similar-
ity among CTI data to make the countermeasures effective. However,
human tasks in using CTI are difficult, because CTI is inherently not
human friendly and a large amount of CTI data is provided. Hence, we
have to spend a lot of time before taking countermeasures. To solve this
problem, we developed an analysis tool which can pick up and visualize
a useful subset of CTI information as a graph, together with attributes
of the organization, to help human judgment. By using graph structure,
the relevancy to the organization and the similarity among CTI data
are revealed at a glance. Moreover, the tool enables the reconciliation
of CTI data, i.e. adding new relationships between them, to store the
result of the analysis for later use. This helps us to take sophisticated
countermeasures.
1 Introduction
1.1 Background
Cyber-attacks range over various area such as Web services, IoT and Cryptocur-
rency. Moreover, there are a lot of ways to attack. Recently, a time period from
beginning of a cyber-attack to achieving its goal, has been shortened. Hence,
preliminary measures and countermeasures that focus on prevention of dam-
age diffusion are necessary. To protect an organization such as infrastructure
facilities and government from cyber-attacks, we have to collect the information
about cyber-attacks in the world such as what kind of attacks are happened,
how it is performed and what kind of information are targeted. If we get a valid
information quickly for organization, we can make the countermeasures effective.
https://doi.org/10.1007/978-3-030-15035-8_65
674 Y. Kambara et al.
Cyber Threat Intelligence (CTI) is evidence-based knowledge including con-

text of cyber-attacks, mechanism, indicators that are used in attacks and coun-
termeasures. Structured Threat Information eXpression (STIX) [4], a language
and serialization format, and Trusted Automated eXchange of Indicator Infor-
mation (TAXII) [5], an exchange protocol for STIX, are developed to share CTI.
Thus, sharing the information of cyber-attacks are underway in the world. CTI
is classified into the following two categories:
• CTI with the emphasis on immediacy, composed of Indicator of Compromise

(IOC) such as an adversary’s server information (IP address, Domain name)
and Hash value of malware.
• CTI with the emphasis on result of analysis. This kind of CTI has rich infor-
mation such as an aim of adversary, a system/geography that was targeted
in addition to IOC.
The former is distributed a lot and used in countermeasures of cyber-attacks

such as block of the communication. The latter has various ways to use such as
log analysis.
1.2 Issue and Purpose
Today, a lot of cyber-attacks are happened in the world and the amount of
CTI data increases in proportion. CTI data enables us to detect cyber-attacks
and prevent diffusion of damage. However, it is difficult for an organization
to find the useful CTI data from a lot of CTI, because effective CTI data for
the organization is very few. We have to extract the CTI data related to the
organization from massive CTI to utilize it. Also, to determine a countermeasure,
we have to analyze cyber-attacks. There is a possibility that a cyber-attack which
seems to be new at first glance is similar to existing attacks. In that case, the
past measures may be applied to the new attack. Hence, We need a mechanism
that reveals whether the past measures can be applied to the new attack by
judging a similarity between the past CTI data and current CTI data.
Then, we paid attention to attributes of the organization. It becomes possible
to extract the related CTI from a large amount of CTI by using this information.
We developed an analysis tool which can pick up and visualize a useful CTI
information and the information of organization as a graph. This tool enables us
to analyze the relevancy between CTI and the organization and reconciliation.
In this paper, we explain the method of efficiently analysis of CTI.
2 Data Set
2.1 Cyber Threat Intelligence
According to Gartner [2], Cyber Threat Intelligence (CTI) is defined as

“evidence-based knowledge of cyber-attacks - including context, mechanisms,
indicators, implications and actionable advice - about an existing or emerging
Developing the Analysis Tool of Cyber-Attacks 675
menace or hazard to IT or information assets. It can be used to inform deci-

sions regarding the subject’s response to that menace or hazard”. For example,
IP address of a command and control server (C&C server) is a part of CTI. In
addition, adversary’s motivation, technique, and attack target are also a part of
CTI. CTI is provided for free from services such as AlienVault’s Open Threat
Exchange (OTX) [1] or NC4’s HAIL A TAXII (HAT) [3]. Also security vendors
such as Symantec and FireEye release their analytical results for fee. In our
research, we use STIX as a CTI format. STIX, the standard format to share
the CTI between organizations, enables us to exchange not only IP address or
domain information but also adversary’s motivation, technique and attack tar-
get. STIX has the following information:
Observables:
The information about evidence of cyber-attack such as adversary’s server
information (IP addresses, Domain names and URLs) and a malware infor-
mation (hash values, sizes and names). Observables are used synonymously
with IOC.
TTPs:
Information about an adversary including methods of a cyber-attack and
targeted systems.
Threat Actors:
Adversary’s motivations, intentions and information about people/
organization that were involved in a cyber-attack.
Courses of Action:
The information about countermeasures to prevent from threat. There are a
lot of information such as circumstances of countermeasures, influence of the
organization, and effectiveness of countermeasures.
There are many methods to get STIX data, for example, TAXII protocol,
which is used exclusively for STIX. The protocol was standardized in 2014.
2.2 Real Data of CTI
According to the official documents of STIX, CTI data written in STIX can
describe information including the adversary’s attack methods, the motivation,
the purpose, the name of threat group and how to deal with the cyber-attack.
We investigated what types of information we can acquire by downloading CTI
data from OTX, HAT and one of the security vendors. Table 1 shows the list of
the data we acquired.
Most of the CTI data we actually acquired is an Observables type. Also,
there exists a little information that is available only for specific service. For
example, the YARA rule used for malware detection can be acquired from OTX
and the port number used in illegal access can be acquired from HAT. In addition
to the Observables, OTX is providing the industry and geography information
of the targeted organization that received damage (TTPs) and threat group
data (Threat Actors). The data of methods and purposes of cyber-attacks were
acquired only from the security vendor’s CTI, but there are no data of the coping
method of cyber-attacks (Courses of Action).
Table 1. Information that can be acquired from CTI
3 Using CTI Data to Protect Organizations
Ideally, an organization on defense side has to check all of the provided CTI data
whether their corresponding threats are appropriately dealt with or not. How-
ever, a large amount of CTI data exists in the Internet and CTI data valuable to
a given organization is rare. Therefore, it takes a long time for an organization
to find CTI data related to directions or operations of its business. To suppress
the damage of cyber-attacks, we have to extract CTI data related to the orga-
nization from a large amount of data and handle quickly CTI data with high
priority. Then, we use the information of organization to reveal a relationship
between CTI data and the organization.
3.1 Grasping Relations Between CTI and Organization
Threat levels of cyber-attacks are different for each organization. For example,
let’s assume there are two adversaries shown in Table 2, and three organizations
shown in Table 3.
Table 2. Adversary’s information
Target geography Target information

CTI1 Japan, United Kingdom Medical information
CTI2 China Government information
Table 3. Organization information
Base region Possession information

Company A Japan Financial data
Company B United Kingdom Medical information
Company C China Government information
For Company A, we can see CTI1 has a high relevance, because the company
has a medical information and located in Japan, and organizations with such
attributes are targeted by CTI1. On the other hand, company A is not targeted
by the adversary of CTI2. Hence, company A has to cope with the CTI1 with
high priority. Company C has to cope with the adversary of CTI2 likewise. In
this way, similarities between adversary’s purpose and organization become an
indicator to attach a priority to several CTI data. We suppose that organization
information has following items:
Global information of active area:

The information about base geographies of the organization and activities for
each base.
Global information of business activity:
The information of activity fields and type of business operations.
Inside information:
Internal material of the organization, information on business and etc.
We sort out these pieces of information and connect them with CTI data.
Table 4 shows corresponding items of CTI data and the information of organiza-
tion. In modern days, these pieces of information are available in the Internet,
which are used by adversaries to attack organizations.
3.2 Grasping the Situation of Countermeasure
Understanding whether the organization has taken measures quickly against

cyber-attacks or not is necessary to cope with cyber-attacks in quick motion.
There are a lot of items that should be checked such as whether it is possible
to detect a malware. We also have to investigate whether the cyber-attack had
already come. We actually obtained cyber-attacks information including hash
value of malware, IP address/Domain name of C&C server and email informa-
tion such as address and title. If we use this information of CTI data and the
Table 4. Corresponding items of CTI data and organization information
information from security equipment, we can judge whether it is necessary to

take the action, and also know circumstances of countermeasure.
4 Analysis Tool of CTI
In this research, we developed an analysis tool of CTI which has the following
functions:
• Extract the useful CTI data from a large amount of CTI

• Visualize the relevancy between CTI data and the organization
• Visualize countermeasure circumstances
• Perform reconciliation to CTI data with GUI and graphical structure
These functions enable us to analyze and take a measure smoothly. In this

section, we explain the flow of handling of cyber-attacks.
4.1 Extracting the Related CTI
It is difficult for the organization to find the useful CTI from a large amount of
CTI. We extract the CTI data which is related to the organization by using the
attribute of it. Then, we calculate the priority of CTI data. Calculating method
is the following. We use “target geography” and “target information” from CTI
and the information of organization. We define Iorg as the set of the attribute
of the organization, Icti as the set of the information which are targeted in the
CTI, Gorg as the set of the geography which locates the organization and Gcti
as the set of targeted geography in the CTI. Then, the priority is defined below.
num (Iorg ∩ Icti ) num (Gorg ∩ Gcti )

priority = +
num (Icti ) num (Gcti )
The possibility of being targeted fluctuates whether the cyber-attack is indis-

criminate or not. Therefore, we normalize the priority so that the priority is low-
ered in the case of an indiscriminate attack. We are not using other information
this time, although we are considering improvement to use it.
Fig. 1. Extract CTI data
In Fig. 1, there are 62 CTI data (blue frame in left). After selection by the
information of organization, the number of CTI is reduced to 4 (blue frame in
right). These remained CTI are related to the organization. To do this, security
operator of the organization can select the CTI which has high priority for them,
and proceed handling of the CTI.
4.2 Visualization of the Relationship Between CTI and the

Organization
Visualizing the attribute of the organization and CTI data enables us to under-
stand the relationship between these two. Figure 2 is an image of the relationship
between attribute of the organization and CTI data. CTI data is in the green
frame. The attribute of the organization is in the red frame. The organization
has to cope with Observable (Fig. 2, right), for example, access blocking. Figure 2
shows the cyber-attack that targets 5 countries (bottom left) and various infor-
mation (orange node) such as Financial information. The organization (in the
red frame) based in 3 locations (Japan, China and United States) and has the
one of the information that targeted by the adversary. Therefore, we can know
that the adversary targets this organization and the organization has to take
a measure. By combining CTI data with the information of organization, we
can understand the relationship between the organization and CTI data at a
glance. In other words, we will know the threat which is the most dangerous in
Fig. 2. CTI data and the attribute of the organization
Fig. 3. CTI witch has high (low) priority for the organization
a short time. Figure 3 shows high priority CTI data and low priority CTI data.
The attribute of the organization and CTI data on the right side has 4 com-
mon nodes. On the other hand, CTI data on the left side has 1 common node.
This indicates that the strength of the relationship between the organization
and right CTI is higher than the other. Hence, we have to cope with CTI on
the right side. Moreover, the red line shows the relevancy among CTI and we
can judge the similarity. If there are two or more CTI, it is possible to apply
the same countermeasure to these CTI. Hence, to judge the similarity enables
to determine countermeasure rapidly.
4.3 Visualization of Countermeasures
To hold on the circumstance of countermeasures speedily, we attached condition

of the organization to Observables which are received from CTI data (green
frame of Fig. 4). We categorized the nodes with colors by whether it already
dealt with or not, for example, in Fig. 4, the color of blocked IP address nodes
is black and the color of other nodes is default (green). There are more items
including Hash Value, Domain name and URL in addition to IP Address to
estimate the circumstance of countermeasure against a threat at a glance.
Fig. 4. Status of CTI data
4.4 Reconciliation of CTI
It is difficult to judge the similarity of CTI by comparing simply the pieces of

data. We made estimating similarity easier by using graphs. Also, reconciliation,
attaching the new relationship between graphs, is possible to store the knowledge
gained thorough analysis. For example, we supposing these situations:
• Putting different hash values which do the same behavior together. This helps
to recognize the malware family.
• Putting different CTI which caused by the same adversary. This helps to
grasp tendency of cyber-attacks together.
Analyzing the fuzzy relation between plural CTI is possible by storing the result
of analysis by human. Also, this helps us to take sophisticated countermeasures.
5 Conclusion
Organization information enables us to cope with cyber-attacks in the order of
descending priorities. We can also extract an useful data which is related to an
organization from a large amount of CTI data. It is easy to check countermeasure
circumstance of CTI data and also easy to analyze a similarity among them by
visualizing it and the information of organization. Future works include verifying
usefulness of our tool and improving it to make a countermeasures more effective.
References
1. AlienVault, Open Threat Exchange (OTX). https://www.alienvault.com/open-
threat-exchange
2. Gartner, Reviews for Security Threat Intelligence Products and Services. https://
www.gartner.com/reviews/market/security-threat-intelligence-services
3. HAIL A TAXII. http://hailataxii.com/
4. OASIS, Introduction to STIX. https://oasis-open.github.io/cti-documentation/
stix/intro
5. OASIS, Introduction to TAXII. https://oasis-open.github.io/cti-documentation/
taxii/intro
6. Masuoka, R., Satomi, T., Yamada, K.: Strategies for semantics-rich machine-
understandable CTI - human - machine cybersecurity collaboration. In: Borderless
Cyber Conference and FIRST Technical Symposium (2017)
7. Katayama, Y., et al.: Use of cyber threat intelligence for organizations. In: Japanese,
Symposium on Cryptography and Information Security (SCIS) (2017)
8. Kambara, Y., et al.: How to set the expiration date of cyber threat intelligence. In:
Japanese, Symposium on Cryptography and Information Security (SCIS) (2017)
Proposal of Ad-Hoc Secure Device
Pairing Method Using Similarity Between
Marker Movement and Acceleration
Makoto Nagatomo1 , Kentaro Aburada2 , Hisaaki Yamaba2 , Naonobu Okazaki2 ,

and Mirang Park1(B)
1
Kanagawa Institute of Technology, 1030 Shimo-Ogino,
Atsugi, Kanagawa 243-0292, Japan
mirang@nw.kanagawa-it.ac.jp
2
University of Miyazaki, 1-1 Gakuen-Kibanadai-Nishi,
Miyazaki, Miyazaki 889-2192, Japan
Abstract. Currently, IoT devices often use wireless technologies to com-

municate with each other. An example is ad-hoc device pairing such as
wireless distribution of meeting materials from a PC to mobile device.
The problem in this case is spoofing by an impersonator. In order to
solve this problem, we proposed secure device pairing method using sim-
ilarity between displacement data of a marker displayed on devices and
acceleration data obtained from devices. We performed experiments to
confirm similarity and whether impersonator can pair with the server.
As a result, we found that similarity was low and unstable. Therefore,
in this paper, we propose an extended method that marker changes at
an interval. These sequence of markers displayed on the device is used to
recognize the device by a camera in addition to calculation of the similar-
ity. We performed 3 experiments to confirm the similarity displacement
data and examined whether an eavesdropper outside the camera range
can perform pairing. Also, we performed pairing of some device together.
As a result, average similarity was higher than that of previous method
and pairing of three devices succeeded at rate of about 70%.
1 Introduction
Recently, along with the advance of wireless technologies, such as Wi-Fi, Blue-
tooth, and near-field communication (NFC), devices such as mobile device and
IoT devices often use these wireless technologies to communicate with each other.
These technologies have privacy vulnerabilities, such as eavesdropping and man-
in-the-middle attack. Hence, it is necessary to establish authenticity between
devices before wireless communication begin. We define secure device pairing as
a process for establish secure wireless communication.
An example of pairing is the wireless distribution of meeting materials from
a PC to mobile devices with the small table (e.g. 3 m × 3 m table) at a meet-
ing. In this pairing, it is necessary to generate a pairing key “on-the-spot”.
https://doi.org/10.1007/978-3-030-15035-8_66
684 M. Nagatomo et al.
The problem in this case is spoofing by a third party outside the room, who can
obtain the meeting materials if secure pairing is not performed between devices
in the room. When Wi-Fi or Bluetooth is used for it, the wireless communica-
tion may be eavesdropped. Therefore, our research goal is to propose the secure
device pairing method, in which is easy to use for users, and it is possible to
perform the pairing in short-range.
Currently, many researches evaluate pairing methods using received signal
strength (RSS) [1,2]. However, it is not possible to perform stable pairing to a
single location because RSS changes significantly due to environmental factors,
such as time and objects.
There is the method which a device having function of emitting visible light
conveys information to the device equipped with a camera [3]. However, the
pairing distance of these methods are short. An alternative is proximity pairing,
in which an infrared camera recognize user’s hand which holds a hand-held device
[4,5]. However, this method requires a special infrared camera. Thus, method
using a regular camera are proposed [6,7]. These methods are tracking methods
of objects with accelerometer. However, these methods is not secure because
camera cannot recognize movements of the objects in detail.
Therefore, we proposed secure device pairing method using similarity between
displacement data of a marker displayed on devices and acceleration data
obtained from devices [8], and performed experiments to confirm similarity and
whether impersonator can pair with the PC. However, we founded that similarity
was not unstable.
In this paper, we propose extended method which marker changes at an
interval. These sequence of markers displayed on the device is used to distin-
guish a device from some devices by a camera in addition to calculation of the
similarity. In addition, we perform three experiments to confirm the similarity
displacement data and acceleration data, and whether an eavesdropper outside
the camera range can perform pairing and perform three device pairing together.
In Sect. 2, we review related works. Section 3 gives a system model and pair-
ing procedure of proposed method. In Sect. 4, we evaluate the accuracy of the
proposed method. Finally, Sect. 5 concludes this research work.
2 Related Work
Amigo [1] is a RSS-based method for pairing a PC and a device. This method
perform identification of a device using machine learning for the features of mean
absolute difference of RSS, mean exponential of difference of RSS, and euclidean
difference of RSS vectors. This method can detect an eavesdropper when the
distance between the attacker and the legitimate devices is 3 m. In addition,
device pairing method in a room using RSS from some APs was proposed [2].
Machine learning is used for the features of mean absolute difference of RSS from
beacon frames in 2.4-GHz band and 5-GHz band obtained from shared APs and
all APs of two devices. This method judge whether devices is in the same room
of the 10 to 15 m2 . An accuracy was higher than 96%. However, RSS varies easily
Proposal of Ad-Hoc Secure Device Pairing Method 685
due to environmental factors, such as time and objects. Thus, it is difficult to

perform stable pairing in this method.
Saxena et al. [3] proposed a device pairing method between a device equipped
with camera and a device equipped with a LED light. This method generates
Diffie-Hellman (DH) key [9] between two devices, the LED device transmits the
hash value of the DH key to the camera device by flashing the LED light in
order to confirm whether the LED device exists in front of the camera device.
However, pairing distance is short (about 30 cm).
Therefore, method which perform pairing between a PC having touch-screen
and smartphone equipped with accelerometer using an infrared camera (Kinect)
is proposed [4,5]. The PC associates the user’s touch and smartphone by depth
image obtained from an infrared camera. This method is not secure because
cannot detect the inclination of the device.
Moving objects tracking methods by a regular camera are proposed [6,7].
The camera recognizes the device equipped with accelerometer using Normalized
Cross-Correlation (NCC). As a result, three people are identified, but these
method do not assume impersonator outside range of camera.
Therefore, we previously proposed secure device pairing method between a
PC equipped with camera and devices equipped with accelerometer [8]. In this
method, the camera first recognizes the displacement data of marker displayed
on devices. At the same time, devices obtain its acceleration data. Second, this
method transforms the displacement data and acceleration data into velocity
data. Finally, this method calculates similarity between the two velocity data.
Pairing completes when the similarity is over a certain value. We performed two
experiments on confirmation similarity and whether impersonator can pair with
the PC, we found similarity was unstable.
3 Proposal of Secure Device Pairing Method Using

Marker Sequence
Figure 1 shows the system model of our proposed method. The components of
this model are as follows.
1. Authentication server (S)

This is a server or other PC equipped with a camera, such as a web camera.
The server calculates the similarity between acceleration data obtained wire-
lessly from the device and movement of the marker displayed on the device
obtained with the camera, and calculates the similarity between sequences of
markers obtained from camera and that obtained from devices.
2. Device (D)
This is a device equipped with accelerometer. This device displays the marker
that the camera can recognize easily. This device transmits the 3-axis accel-
eration data and sequence of markers to the server wirelessly at pairing. In
this system model, it is assumed that some devices simultaneously perform
the pairing with the server, hence there exists n devices.
The pairing procedure is as follows.
Step 1: Display marker

D displays a marker. This maker is a simple marker that the camera can
detect from a distance of several meters.
Step 2: Recognize marker
Server S recognizes the marker on the D’s display through the S’s camera.
Step 3: M ove Devices
Users move own devices as the marker reflects on the camera. In this phase,
type of marker changes at an interval.
Step 4: Acquisition of movement, acceleration data and M arker Sequence
D wirelessly transmits the set α of acceleration data of x, y and z axis
obtained from the device, types of marker sequence displayed in step 3, and
its acquisition time to S. S creates the marker movement dataset β, which
captures motion in the x and y axes, marker sequence recognized in Step 3,
and its acquisition time. α and β are represented as follows:
α = {(axi , ayi , azi , mα α

i , ti )|i ∈ {1, ..., m}} (1)
β= {(xj , yj , mβj , tβj )|j ∈ {1, ..., n}} (2)
where axi , ayi , azi , mα

i (i ∈ {1, ..., m}) represent acceleration along x, y and
β
z axis, and type of displayed marker obtained at time tα i . xj , yj , mj (j ∈
{1, ..., n}) represent the coordinates of the marker on the image and its type
of marker at time tβj .
Step 5: Eliminate noise
Server S eliminates the noise and gravitational acceleration in acceleration
data using Fast Fourier Transform (FFT).
Step 6: Convert data to velocity data
The S differentiates the motion data, and integrates the acceleration data to
obtain the velocity data. In addition, we use function of spline interpolation
to displacement data in β to adjust the number of data to the number of
acceleration data. Each velocity datum α and β is represented as follows:
α = {(vix , viy , viz , mα α

i , ti )|i ∈ {1, ..., m}}, (3)
β = {(xi , yi , mβi , tβi )|i ∈ {1, ..., m}}. (4)
Step 7: N ormalize data

It is impossible to calculate the similarity between α and β because the units
of two data are different. Therefore, we perform normalization for each axis
by setting magnitude of a vector to 1. This vector is the data arranged in time
order for each axis (the dimension of the vector is m). Each α̃, β̃ obtained by
normalizing α and β is as follows:
α̃ = {(ṽix , ṽiy , ṽiz , mα α

i , ti )|i ∈ {1, ..., m}}, (5)
β̃ = {(x̃i , ỹi , mβi , tβi )|i ∈ {1, ..., m}}, (6)
where ṽiw , ũi (w ∈ {x, y, z}, u ∈ {x, y}, i ∈ {1, ..., m}) are calculated as follows:
m
m

i
ṽw = viw / (vkw )2 , ũi = ui / u2
k, (7)
k=1 k=1
and ṽiw and ũi take values from −1 to 1.

Step 8: Calculate Similarity of M arker Sequence
Server S calculate the similarity between marker sequences obtained from S
and D in α̃ and β̃.
Step 9: Calculate Similarity between Displacement Data Acceleration Data
Server S calculates the similarity of velocity data in α̃ and β̃. When the simi-
larities of marker sequences and velocity datum are higher than the threshold
θ1 and θ2 , the pairing is accepted. S and D then generate a common key by
DH key exchange.
Benefits: This method has longer pairing distance than QR code method
because a marker of this method can be more simple than that of QR code
method. Hence, one-to-many device pairing is possible by reading markers on
the displays of several devices.
In our proposed method, server S calculates the sequences of markers
obtained from camera and device in step 8. S calculates the similarity ms as
follows:
β
number that mα
i = mi in α̃ and β̃
ms = (8)
number of set α̃ or β̃(m)
In addition, server S calculates the average of similarity of x and y axes as the

similarity between α̃ and β̃. We used four methods for calculating the similarity
in [8], and found that the correlation coefficient is useful in our proposed method.
Thus we use the correlation coefficient as calculation of similarity. Note that we
do not consider calculation of similarity using z axis data of acceleration and
inclination of the marker at this time. The similarity is calculated as following
equation.
m
(ṽiw − ṽ w )(w̃i − w̃)
rw = m i=1 m , (9)
( i=1 (ṽiw − ṽ w )2 )( i=1 (w̃i − w̃)2 )
m m
where w ∈ {x, y}, ṽ w = k=1 ṽkw , and w̃ = k=1 w̃k . rw takes a value from −1
to 1. It can be judged that there is a positive correlation when rw is positive.
When rw is negative, there is a negative correlation.
4 Evaluation Experiment and Consideration

We implemented a prototype for evaluation experiment of the proposed method
in Sect. 3. The server was a laptop PC (DELL MI11C-6WHBR), and the device
was mobile device (Nexus 5X and Nexus 7). We implemented using python3 for
the PC program and using java in Android Studio for the device. The camera
equipped with the laptop PC is used for camera and accelerometer that were built
into the laptop and smartphone, respectively. Marker recognition was performed
by ArUco [10], which is a library in OpenCV.
Figure 2(a) shows the device application. The application starts the wire-
less communication with PC’s application when the user starts the application.
When the user presses the pairing button, the application starts to obtain the
acceleration data of x, y and z axis and the obtained time. The application trans-
mits the set of the data when the button is pressed again. Marker changes at
0.5 s interval.
Figure 2(b) shows the PC application acquiring the device image. The video
obtained from the camera is displayed when the application starts. We set the
size of the video as 640 × 360 pixels. When the device user starts the device
application, it initiates wireless communication with PC application. The camera
gets the 4 coordinates of the marker’s corners. Hence, the application can obtains
the average coordinate of the 4 coordinates as displacement data.
We show an example of velocities calculated from the acceleration and dis-
placement data in Fig. 3(a) and (b). The horizontal axis represents the time from
the initiation of pairing. vel.X and vel.Y represent the obtained data along x and
y axis. We found that the shapes of graphs are almost the same.
4.1 Experiment for Similarity

In the proposed method in Sect. 3, the similarity probably varies according to the
device motion. Therefore, we performed a confirmation experiment to evaluate
this. We recruited 16 students belonging to Kanagawa Institute of Technology.
The procedure of this experiment is as follows.
1. The student holds the device 1.0 m away from the camera and launches the
device application.
2. The student holds the device for the laptop camera. The student moves the
device along the shape of a circle or ∞ symbol for about 5 s.
(a) Application of device (b) Application of PC
Fig. 2. Implemented application
(a) Velocity data converted from displacement (b) Velocity data converted from acceleration
data data
Fig. 3. Normalized velocity data
3. The PC application calculates the similarity between the data obtained in

step 3.
4. The student repeats steps 1–3 five times.
In this experimental environment, the maximum distance at which the cam-

era could detect the marker was 1.0 m. Thus, we performed this experiment at
1.0 m.
Table 1 shows the average similarity of marker sequences. Similarity using
Nexus 7 (tablet) was higher than that using Nexus 5X (smartphone). Thus, we
can find that the larger the display is, the larger the similarity is. Table 2 shows
the average similarity of velocity data. The average similarity was larger than
that in previous method [8] over 0.1. However, proposed method is unstable
because standard deviation was large.
4.2 Experiment for Illegal Pairing
There is a possibility that the third party who is outside the range of the camera
imitates the movement of the legitimate user. If the similarity has variability,
Table 1. Average and standard deviation of similarity of marker sequence
Device, motion Nexus 5X, circle Nexus 5X, ∞ Nexus 7, circle Nexus 7, ∞ Average
Average similarity 0.717 0.688 0.855 0.854 0.778
Standard deviation 0.248 0.217 0.213 0.199 0.232
Table 2. Average and standard deviation of similarity of velocity data
Device, motion Nexus 5X, circle Nexus 5X, ∞ Nexus 7, circle Nexus 7, ∞ Average
Average similarity 0.571 0.575 0.586 0.619 0.588
Standard deviation 0.265 0.223 0.292 0.197 0.197
that of the impersonating device can be higher than that of legitimate device.
Therefore, we perform the confirmation experiment for considering it. Subjects
are 12 university students. We performed this experiment as following procedure.
1. Legitimate user holds the device 1.0 m away from the PC’s camera. Both start
the device’s application.
2. The user holds the device for the camera. After that, user moves the device
along the shape of circle or ∞ symbol for 5 s. Simultaneously, impersonator
imitates the movement of the user’s device.
3. The PC calculates the similarity between the displacement data obtained in
step 2 and the acceleration data which is transmitted from both user’s device
and impersonator’s device in step 2.
4. We repeat 1–3 five times.
Table 3 shows the average similarities of marker sequence. The difference of

the similarity between legitimate user and impersonator was very large.
Table 4 shows the average similarities of velocity data between legitimate
users and impersonators. The result shows that the average similarities of legit-
imate user are higher than to that of impersonator over 2.0. Previous method
[8] has difference of average similarity under 2.0. Hence, the proposed method
can distinguish legitimate user from impersonator more than previous method.
In addition, it is possible to determine a threshold of similarity, but the stan-
dard deviations of the similarity was large. We found that this pairing method
is unstable in this regard.
4.3 Pairing of Several Devices at the Same Time

We perform experiment whether several devices can be performed pairing
together. Subjects are 12 university students. Devices used in this experiment
are two Nexus 5Xs and a Nexus 7. Experimental procedure is as follows.
1. We make the team consisted of 3 people.
2. The 3 people launch devices’ application, and move devices on the motion of
circle or ∞.
Table 3. Average similarity of marker sequence of legitimate user and impersonator
Legitimate user Impersonator

Average similarity 0.747 0.016
Standard deviation 0.276 0.030
Table 4. Average similarity of velocity data of legitimate user and impersonator
Legitimate user Impersonator

Average similarity 0.557 0.330
Standard deviation 0.294 0.352
3. The PC application calculates the similarity for each devices.

4. We repeat steps 2 and 3 five times.
As a result, Three devices were distinguished correctly at rate of 70% by

comparing marker sequences of devices. In addition, average similarity of velocity
data and its standard deviation were 0.461 and 0.299. We can find that the
average similarity on this experiment is lower than that on first experiment for
similarity.
5 Conclusion
In this paper, we proposed an extended method for pairing between a server

equipped with a camera and a device equipped with the accelerometer using
marker. In addition, we performed the three experiments on confirmation of
similarity, whether several devices can distinguish and illegal pairing can be per-
formed. The result of the first experiment showed that the similarity is large
as the size of device’s display is large. Also, the average of similarity in the
proposed method was higher than that of previous method. The result of the
second experiment showed that we can separate the legitimate user and imper-
sonator by the average of the similarity more than previous method. However,
the stable pairing cannot be performed because the variation of the similarity
was large. The result of the third experiment showed that some devices can be
performed the pairing together, but the rate that the proposed method can dis-
tinguish three devices was 70%. The future work are reduction of the variation
for the similarity and raise of distinguished rate of some devices, and to perform
experiment using camera having high resolution.
References
1. Varshavsky, A., Scannell, A., LaMarca, A., de Lara, E.: Amigo: proximity-based
authentication of mobile devices. In: Proceedings of the 9th International Confer-
ence on Ubiquitous Computing (UbiComp 2007), pp. 253–270 (2007)
2. Agata, Y., Hong, J., Ohtsuki, T.: Room-level proximity detection based on RSS of
dual-band Wi-Fi signals. In: Proceedings of 2016 IEEE International Conference
on Communications (ICC) (2016)
3. Saxena, N., Ekberg, J.E., Kostianien, K.: Secure device pairing based on a visual
channel: design and usability study. IEEE Trans. Inf. Forensics Secur. 6(1), 28–38
(2010)
4. Rofouei, M., Wilson, A.D., Brush, A.J.B., Tansley, S.: Your phone or mine? Fus-
ing body, touch and device sensing for multi-user device-display interaction. In:
Proceedings of the SIGCHI Conference on Human Factors in Computing Systems,
pp. 149–158 (2012). Camera View, Intelligent Robots and Systems, pp. 3872–3877
(2008)
5. Wilson, A.D., Benko, H.: CrossMotion: fusing device and image motion for user
identification, tracking and device association. In: Proceedings of the 16th Inter-
national Conference on Multimodal Interaction, pp. 216–223 (2014)
6. Maruhashi, N., Terada, T., Tsukamoto, M.: A method for identification of moving
objects integrative use of a camera and accelerometers. In: Proceedings of the 27th
Annual ACM Symposium on Applied Computing, pp. 1–6 (2012)
7. Osamu, S., Kagami, S., Hashimoto, K.: Identifying a moving object with an
accelerometer in a camera view. In: Proceedings of the 16th International Con-
ference on Multimodal Interaction, pp. 216–223 (2014)
8. Nagatomo, M., Aburada, K., Okazaki, N., Park, M.: Proposal and evaluation of
secure device pairing method with camera and accelerometer. In: Proceedings of
2018 Sixth International Symposium on Computing and Networking Workshops
(CANDARW), pp. 310–315 (2018)
9. Diffie, W., Hellman, M.E.: New directions in cryptography. IEEE Trans. Inf. The-
ory 22(6), 644–654 (1976)
10. Aruco. https://docs.opencv.org/3.2.0/d5/dae/tutorial aruco detection.html. Acc-
essed 28 Jan 2019
Introduction of Fingerspelling
for Realizing a User Authentication
Method Using s-EMG
Hisaaki Yamaba1(B) , Shimpei Inotani1 , Shotaro Usuzaki1 , Kayoko Takatsuka1 ,

Kentaro Aburada1 , Tetsuro Katayama1 , Mirang Park2 , and Naonobu Okazaki1
1
University of Miyazaki, Miyazaki, Japan
yamaba@cs.miyazaki-u.ac.jp
2
Kanagawa Institute of Technology, Atsugi, Japan
Abstract. At the present time, mobile devices such as tablet-type PCs

and smart phones have widely penetrated into our daily lives. Therefore,
an authentication method that prevents shoulder surfing is needed. We
are investigating a new user authentication method for mobile devices
that uses surface electromyogram (s-EMG) signals, not screen touch-
ing. The s-EMG signals, which are detected over the skin surface, are
generated by the electrical activity of muscle fibers during contraction.
Muscle movement can be differentiated by analyzing the s-EMG. Tak-
ing advantage of the characteristics, we proposed a method that uses a
list of gestures as a password in the previous study. In order to realize
this method, we have to prepare many gestures that are used to com-
pose passwords. In this paper, we adopted fingerspelling as candidates
of such gestures. We measured s-EMG signals of manual kana of The
Japanese Sign Language syllabary and examined their ability for the use
of passwords of the authentication method.
1 Introduction
This paper presents an introduction of fingerspelling as candidates of gestures
used in the user authentication method for mobile devices by using surface elec-
tromyogram (s-EMG) signals, not screen touching.
An authentication method that prevents shoulder surfing, which is the direct
observation of a users personal information such as passwords, comes to be
important. At the present time, mobile devices such as tablet type PCs and
smartphones have widely penetrated into our daily lives. So, authentication oper-
ations on mobile devices are performed in many public places and we have to
ensure that no one can view our passwords. However, it is easy for people who
stand near such a mobile device user to see login operations and obtain the users
authentication information. And also, it is not easy to hide mobile devices from
attackers during login operations because users have to see the touch screens of
their mobile devices, which do not have keyboards, to input authentication infor-
mation. On using a touchscreen, users of a mobile device input their authentica-
tion information through simple or multi-touch gestures. These gestures include,
https://doi.org/10.1007/978-3-030-15035-8_67
694 H. Yamaba et al.
for example, designating his/her passcode from displayed numbers, selecting reg-
istered pictures or icons from a set of pictures, or tracing a registered one-stroke
sketch on the screen. The user has to see the touch screen during his/her login
operation; strangers around them also can see the screen.
To prevent this kind of attack, biometrics authentication methods, which use
metrics related to human characteristics, are expected. In this study, we inves-
tigated application of surface electromyogram (s-EMG) signals for user authen-
tication. s-EMG signals, which are detected over the skin surface, are generated
by the electrical activity of muscle fibers during contraction. These s-EMGs
have been used to control various devices, including artificial limbs and electri-
cal wheelchairs. Muscle movement can be differentiated by analyzing the s-EMG
[1]. Feature extraction is carried out through the analysis of the s-EMGs. The
extracted features are used to differentiate the muscle movement, including hand
gestures.
In the previous researches [2–7], we investigate the prospect of realizing an
authentication method using s-EMGs through a series of experiments. First,
several gestures of the wrist were introduced, and the s-EMG signals generated
for each of the motion patterns were measured [2]. We compared the s-EMG
signal patterns generated by each subject with the patterns generated by other
subjects. As a result, it was found that the patterns of each individual subject are
similar but they differ from those of other subjects. Thus, s-EMGs can confirm
ones identification for authenticating passwords on touchscreen devices. Next,
a method that uses a list of gestures as a password was proposed [3,4]. And
also, a series of experiments was carried out to investigate the performance of
the method extracting feature values from s-EMG signals adopted in [5] and the
methods to identify gestures using dynamic data warping (DTW) [6] and support
vector machines (SVM) [7]. The results showed that the methods are promising
to identify gestures using extracted feature values from s-EMG signals.
In this paper, we adopted figerspelling as candidates of gestures used in this
authentication method. To prepare enough number of gestures that are needed
to compose a password, which is a list of gestures, The Japanese Sign Language
syllabary were adopted and examined their characteristics.
Fig. 1. A sample of an s-EMG signal

Introduction of Fingerspelling for a User Authentication Method 695
2 Characteristics of Authentication Method for Mobile

Devices
It is considered that user authentication of mobile devices has two
characteristics [2].
One is that an authentication operation often takes place around strangers.
An authentication operation has to be performed when a user wants to start
using their mobile devices. Therefore, strangers around the user can possibly see
the user’s unlock actions. Some of these strangers may scheme to steal informa-
tion for authentication such as passwords.
The other characteristic is that user authentication of mobile devices is almost
always performed on a touchscreen. Since many of current mobile devices do not
have hardware keyboards, it is not easy to input long character based passwords
into such mobile devices. When users want to unlock mobile touchscreen devices,
they input passwords or personal identification numbers (PINs) by tapping num-
bers or characters displayed on the touchscreen. Naturally, users have to look at
their touchscreens while unlocking their devices, strangers around them also can
easily see the unlock actions. Besides, the user moves only one finger in many
cases. So, it becomes very easy for thieves to steal passwords or PINs.
Fig. 2. A list of gestures used as a password
To prevent shoulder-surfing attacks, many studies have been conducted. The

secret tap method [8] introduces a shift value to avoid revealing pass-icons. The
user may tap other icons in the shift position on the touchscreen, as indicated
by a shift value, to unlock the device. By keeping the shift value secret, people
around the user cannot know the ture pass-icons, although they can still watch
the tapping operation. The rhythm authentication method [9] relieves the user
from looking at the touchscreen when unlocking the device. In this method, the
user taps the rhythm of his or her favorite music on the touchscreen. The pattern
of tapping is used as the password. In this situation, the users can unlock their
devices while keeping them in their pockets or bags, and the people around them
cannot see the tap operations that contain the authentication information.
3 User Authentication Using s-EMG

The s-EMG signals (Fig. 1) are generated by the electrical activity of muscle
fibers during contraction and are detected over the skin surface [2]. Muscle move-
ment can be differentiated by analyzing the s-EMG.
In the previous research, the method of user authentication by using s-EMGs
that do not require looking at a touchscreen was proposed [3,4]. The s-EMG
signals are measured, and the feature values of the measured raw signals are
extracted. We estimate gestures made by a user from the extracted features. In
this study, combinations of the gestures are converted into a code for authenti-
cation. These combinations are inputted into the mobile device and used as a
password for user authentication.
1. At first, pass-gesture registration is carried out. A user selects a list of gestures

that is used as a pass-gesture (Fig. 2(a)).
2. The user measures s-EMG of each gesture, extracts their feature values, and
register the values into his mobile device (Fig. 2(b)).
3. When the user tries to unlock the mobile device, the user reproduces his
pass-gesture and measures the s-EMG.
4. The measured signals are sent to his mobile device.
5. The device analyzes the signals and extracts the feature values.
6. The values are compared with the registered values.
7. If they match, the user authentication will succeed (Fig. 2(c)).
8. On the other hand, an illegal user authentication will fail because a list of
signals given by someone who stole the device (Fig. 2(d)) will not be similar
with the registered one.
Adopting s-EMG signals for authentication of mobile devices has three advan-
tages. First, the user does not have to look at his/her device. Since the user can
make a gesture that is used as a password on a device inside a pocket or in a
bag, it is expected that the authentication information can be concealed. No
one can see what gesture is made. Next, it is expected that if another person
reproduces a sequence of gestures that a user has made, the authentication will
not be successful, because the extracted features from the s-EMG signals are
usually not the same between two people. And then, a user can change the list
of gestures in our method. This is the advantages of our method against other
biometrics based methods such as fingerprints, an iris, and so on. When authen-
tication information, a fingerprint or an iris, come out, the user can’t use them
because he/she can’t change his/her fingerprint or iris. But the user can arrange
his/her gesture list again and use the new gesture list.
4 Collecting Gestures for the Authentication Method

Using s-EMG
In order to realize the proposed authentication method using s-EMG, we have
to prepare many gestures that represent characters of passwords. Such gestures
have to be easy to tell each of them from others.
We are planning to adopt such gestures referring to fingerspelling. Finger-
spelling is the representation of the letters using hands. The set of manual signs is
Fig. 3. The Japanese Sign syllabary. (https://upload.wikimedia.org/wikipedia/comm

ons/d/dc/JSL-AIUEO.jpg)
Fig. 4. Measuring an s-EMG signal
used to help the communication using sign languages. For example, proper nouns
such as names of persons are represented by fingerspelling. Sign languages such
as American Sign Language, French Sign Language, British Sign Language, and
so on, have there own manual alphabet.
We adopted the Japanese Sign syllabary (see Fig. 3) as the candidates of
gestures used in our authentication method. They are called yumimoji, which
means “finger letters”. Comparing with manual alphabets representing Latin
alphabet, there are larger numbers of manual kana in yumimoji, 46 letters.
By adopting gestures referring figerspelling, a password that is made up of
gestures corresponds to a string that is made up of letters. It is expected that
this helps users to remember their passwords.
However, some yubimoji may not generate distinct s-EMG signals. To select
yumimoji that are suitable for the use of this study, we examined their perfor-
mances as pass-gesture by measuring their s-EMG signals.
5 Experiments
5.1 Purpose
A series of experiments was carried out to investigate the prospect of the authen-
tication method using yubimoji. Specifically, we investigated whether the mea-
sured s-EMG signals of one experimental subject were similar and whether the
signals of different letters were different from each other.
5.2 Conditions
In these experiments, the set of DL-3100 and DL-141 (S&M Inc.) that was an
electromyograph used in the previous researches also used to measure the s-
EMG of each movement pattern in this study (Fig. 4.) Experimental subjects
made gestures by their left hand. Subjects sit on an arm chair and put his
forearm on the armrest. Two electrode sensors were put on the palm side of the
forearm at the positions of 4 cm and 6 cm toward the elbow from the wrist (The
former is called “m1” and the latter is called “m2” in the following sections).
The measured data were stored and analyzed on a PC.
Five students of University of Miyazaki participated as experimental subjects.
And the 46 yubimoji shown in Fig. 3 were examined. However, since a subject
put his forearm on an armrest, we arranged some yubimoji. When we make a
yubimoji like TE, the palm faces forward. But in these experiment, the palm
faces upward.
Before making each yubimoji, subjects clenched their left fists. s-EMG signals
making yubimoji from a clenched fist were measured. The subjects repeated each
yubimoji ten times and their s-EMG signals were recorded. This measurement
was carried out 3 times and 30 signals were obtained for each subject and for
each yubimoji.
Fig. 5. A part of the results of the experiments.

5.3 Results
The number of s-EMG signal data is not so many, but we found that some
yubimoji letters had some caracteristics.
(1) yubimoji that bends one’s wrist forth
The yubimoji of KU, SHI, NI, MI, YO bend one’s wrist forth. Some of their
fingers are extended pointing to the side.
Amplitude of s-EMG signals measured by m2 are large when the yubimoji
is one of this type (Fig. 5(a), (b)).
(2) yubimoji that points downward
The yubimoji of SU, NE, FU, HE also bend one’s wrist forth but their
extended fingers point downward.
Amplitude of s-EMG signals measured by both m1 are as large as (1) and m1
(Fig. 5(d)). Amplitude measured by m2 is also as large as (1) if its middle finger
is extend (SU and NA.) Other (HE and HU) wise they were smaller (Fig. 5(e)).
(3) yubimoji that extend only one finger
The yubimoji of A, I, SE, TA, HI extends only one finger. Also they don’t
bent one’s wrist.
Amplitude of s-EMG signals measured by both m1 and m2 are very slight
(Fig. 5(c)). It seems hard to use these letters as elements of pass-gestures.
(4) yubimoji that extend more than one finger
The yubimoji of U, KE, TE, WA extends more than one finger including
index finger and middle finger. Also they don’t bent one’s wrist.
Amplitude of s-EMG signals measured by m1 are very slight. Amplitude
measured by m2 are larger than those by m1 but smaller than (1) and (2)
(Fig. 5(f)).
From the results of the experiments mentioned above, one’s wrist is bent or
not and some finger is extended or not are expected to be promising clues to
distinction of yubimoji.
6 Conclusion
We investigated a new user authentication method that can prevent shoulder-
surfing attacks in mobile devices. To realize the authentication method using
s-EMG, we examined the characteristics of the yubimoji, the Japanese Sign
Language syllabary, as the candidate of the element of pass-gestures. A series
of experiments was carried out to compare the measured s-EMG signals gener-
ated by making gesutures based on yubimoji. Results of the experiments showed
that which fingers are extended and whether a wrist is bent or not can be used
as characteristics to identify yubimoji from s-EMG signals. We are planning to
examine characteristics of yubimoji using a lot of s-EMG data from many people.
Also, we would like to explore appropriate positions of measuring s-EMG.
Acknowledgements. This work was supported by JSPS KAKENHI Grant Numbers

JP17H01736, JP17K00139, JP17K00186, JP18K11268.
References
1. Tamura, H., Okumura, D., Tanno, K.: A study on motion recognition without FFT
from surface-EMG (in Japanese). IEICE Part D J90–D(9), 2652–2655 (2007)
2. Yamaba, H., Nagatomo, S., Aburada, K., et al.: An authentication method for
mobile devices that is independent of tap-operation on a touchscreen. J Robot.
Netw. Artif. Life 1, 60–63 (2015)
3. Yamaba, H., Kurogi, T., Kubota, S., et al.: An attempt to use a gesture control arm-
band for a user authentication system using surface electromyograms. In: Proceed-
ings of 19th International Symposium on Artificial Life and Robotics, pp. 342–245
(2016)
4. Yamaba, H., Kurogi, T., Kubota, S., et al.: Evaluation of feature values of surface
electromyograms for user authentication on mobile devices. Artif. Life Robot. 22,
108–112 (2017)
5. Yamaba, H., Kurogi, T., Aburada, A., et al.: On applying support vector machines
to a user authentication method using surface electromyogram signals. Artif. Life
Robot. (2017). https://doi.org/10.1007/s10015-017-0404-z
6. Kurogi, T., Yamaba, H., Aburada, A., et al.: A study on a user identification method
using dynamic time warping to realize an authentication system by s-EMG. Adv.
Internet Data Web Technol. (2018). https://doi.org/10.1007/978-3-319-75928-9 82
7. Yamaba, H., Aburada, A., Katayama, T., et al.: Evaluation of user identification
methods for an authentication system using s-EMG. Adv. Netw.-Based Inf. Syst.
(2018). https://doi.org/10.1007/978-3-319-98530-5 64
8. Kita, Y., Okazaki, N., Nishimura, H., et al.: Implementation and evaluation of
shoulder-surfing attack resistant users (in Japanese). IEICE Part D J97–D(12),
1770–1784 (2014)
9. Kita, Y., Kamizato, K., Park, M., et al.: A study of rhythm authentication and
its accuracy using the self-organizing maps (in Japanese). Proc. DICOMO 2014,
1011–1018 (2014)
App-Collusion Detection Using
a Two-Stage Classifier
Md. Faiz Iqbal Faiz(B) , Md. Anwar Hussain, and Ningrinla Marchang
North Eastern Regional Institute of Science and Technology, Itanagar, India

{rs faiz ece15,ah,nm}@nerist.ac.in
Abstract. Various single app analysis tools are developed for checking
Android malwares in the smartphones. All of them are unable to detect
threats caused by more than one app. Android app-collusion is one such
threat. App-collusion is a scenario where two or more apps collaborate
with each other to achieve a malicious goal. This paper presents an app-
roach for detecting collusive app-pairs using machine learning. The pro-
posed approach works in two stages. In the fist stage, we train a base
classifier using a set of benign and malicious applications. In the second
stage, we use the parameter vector from the first stage and a different
classifier to detect collusive app-pairs. We achieve detection rates of 90%
and 87.5% on two sets of colluding app-pairs.
1 Introduction
Android employs permission-based security mechanism. In this mechanism, data
of individual apps are protected from unauthorized access. Each app is installed
with permissions. These permissions are requested at the time of installation.
The permissions protect data of individual apps from other apps. We can say that
the permission-based security mechanism protects private data of users. During
app installation, if the requested permissions are not granted by the user then
the installation is cancelled (for devices running Android versions below 6.0).
But for versions above 6.0, the apps can ask for permissions during run-time.
The users can accept or deny the requested permission.
Individual apps are executed in separate sandboxes. Protecting sensitive
resources from unauthorized access by the third-party apps. The Android operat-
ing system cannot detect if an app accessing a permission-protected resource via
another app has itself requested the permission. An app can have access to sensi-
tive resources which are permission protected and can limit access to its data by
using permissions. The developer of the app can expose the permission-protected
resources advertently or inadvertently. The applications can collaborate with
each other to gain access to permission-protected resources. Android malware
tools are capable of analysing single apps. However, they cannot detect threats
that span multiple apps. The colluding apps have permissions which when com-
bined, can bypass any security rule. The individual apps in the colluding app
set may appear benign and incapable of executing a threat [1,4,6,8,11,12].
https://doi.org/10.1007/978-3-030-15035-8_68
App-Collusion Detection Using a Two-Stage Classifier 703
Android app-collusion is an emerging threat. In a recent report published by

Intel security, more than 5000 versions of 21 mobile apps containing colluding
code have been found. These apps can carry out information theft, faking SMS
messages, and other kinds of threat [15]. Mobile anti-virus tools at present can
detect only single malicious apps. Threats like app-collusion can easily bypass
these tools. Hence, there is an urgent need to develop multi-app analysis tools
that can analyse multiple apps and check whether they are secretly communi-
cating with each other. App-collusion can be exploited by using (a) permissions
of the applications (b) covert or overt channels between applications, to achieve
a malicious goal [2–4,7,9,10,14].
This paper proposes a novel and simple technique to detect colluding app-
pairs. The contributions of the paper are summarized as follows:
• Development of an app-collusion detection system based on machine learning
• Empirical analysis of various permissions which are crucial with regards to
app-collusion
2 Related Work
The idea of app collusion is first explained in [10], in which it is demonstrated
that Android’s permission-based mechanism is incapable of preventing appli-
cation collusion attacks. The authors showed using throughput measurements
of communication channels that even covert channels with low throughput can
leak private data out of smartphone. They posited app collusion is an open prob-
lem for the research community. The authors in [4] proposed XManDroid which
is an extension of Android permission framework to tackle privilege escalation
attacks. It analyses communication links between applications and ensures that
they comply to the desired system policy. The authors in [9] gave the first col-
lusion threat intelligence framework to detect app collusion. The system was
based on permission-based features and used Naive Bayes classifier with infor-
mative prior to calculate collusion threat. They also presented extensive analysis
of the role of permissions in collusion threat estimation. App-collusion detection
using informative logistic regression classifier is presented in [7], in which authors
discuss impact of various permissions.
3 Methodology
In this work, we present an app-collusion detection system for detecting app-
pairs which are colluding. We discuss app-collusion in Sect. 3.1, app-collusion
detection in Sect. 3.2, and finally the system model in Sect. 3.3.
3.1 App-Collusion
App-collusion is a scenario in which two or more apps secretly communicate with
each other and carry out a threat. The threat can be in the form of information
704 Md. F. I. Faiz et al.
theft, financial theft, service misuse, and elevation of privilege [15]. Information
exfiltration is one of the most prominent threat. Information leakage can be
carried out using the Internet feature or external storage of the device.
App-collusion is a multiple-app attacking model. The operations needed to
execute a threat are divided among multiple apps. However, when we look at
a single app, it appears benign. Each app can execute a portion of the given
operations using its permissions. When we look at it collectively, the colluding
app set includes dangerous combination of permissions capable of achieving a
malicious goal.
We give a formal definition of app collusion as in [7]. Assume A to be the set
of all android apps and P be the set of all dangerous permissions. Let a and b
be two applications with permission sets Pa and Pb such that a, b ∈ A and Pa ,
Pb ∈ P .
Now, assume that a performs a sensitive operation to access resource τ by
using permission β such that β ∈ (Pa −Pb ). If f (τ ) is a set of sequence of sensitive
operations on/using τ and if b performs a sequence of operations α ⊆ f (τ ) using
permission γ ∈ Pb . Then, this scenario is called collusion.
3.2 App-Collusion Detection
In this work, the number of apps in the colluding set is limited to two. But this
can be easily extended to more than two apps. We first train a machine learning
classifier using a corpus of benign and malicious applications. The features for
the classifier are critical permissions which are frequently requested by Android
malwares. We call this the base classifier for our detection system. More details
are provided in Sect. 3.3. We assume that threats posed by colluding apps are
no different than those posed by single apps.
After the base classifier is trained with a set of benign and malicious appli-
cations, we extract the learned parameters of each feature. The output of the
base classifier is the parameter vector. Each parameter in the parameter vector
corresponds to a feature of the data samples. The resultant parameter vector is
fed to a final classifier. This classifier takes the learned parameter vector and an
app-pair as input. The output of the classifier is a decision signalling whether
the given app-pair is colluding or non-colluding.
In the next section, we explain in detail our system model and the classifiers
used in it.
3.3 System Model
The app-collusion detection system consists of two main components. The first
component is the base classifier and the second component is the final classifier.
In this model, we use naive Bayes classifier as the base classifier. The base
classifier is trained with a set of benign and malicious applications. The features
we use are the critical permissions. These permissions are frequently requested by
Android malwares and some fall under dangerous category of permissions. We get
the parameters that is calculated by training the base classifier. Each parameter
corresponds to a feature. So, the number of parameters is equal to the number of
features. These parameters are used by the final classifier for detecting collusive
app-pairs. The rationale behind using naive Bayes classifier as the base classifier
and a simple final classifier is that they are not compute-intensive compared
to other machine learning algorithms such as neural networks. We know that
the smartphones have limited amount of computational resources, hence using
low compute-intensive and simple classifier is the best choice. In the following
section, we explain the two main components.
3.3.1 Base Classifier

We represent the training set as X = {xi ; i = 1, 2, . . . , n} where n is the total
number of samples (applications) in the training set. Each app is represented by
a feature (permission) vector xi = (xi1 , . . . , xik ). Each xij is a Bernoulli random
variable which takes the value 1 if permission j is requested by the corresponding
app, 0 otherwise. We assume that the user has no prior knowledge about the
maliciousness of the app and grants permissions requested by the app [7]. We
define a binary classifier, f : X → C where C = {c1 , c2 }. Here, c1 and c2
corresponds to non-malicious class and malicious class. We assume that the
features are independent and identically distributed which implies P (xij |cl ) =
P (xj |cl ) ∀ i. According to naive Bayes classifier, for each new sample xi =
(xi1 , . . . , xik ).
cnb = argmax P (cl ) P (xj |cl ) (1)
cl ∈C j
Where cnb represents the target value output by the naive Bayes classifier.
Here, P (xj |cl ) are the feature probabilities conditioned on the classes. Detailed
explanation on naive Bayes classifier is found in [5]. After training the classifier
with the training set, we get the parameter vector θl = (θ1l , . . . , θkl ) for each class
cl where each parameter: θjl = P (xj |cl ). We use the parameter vector θ in the
final classifier.
3.3.2 Final Classifier

In the final step, we use the final classifier. We use the idea that a colluding
app-pair is no different than a malicious application. We treat each app-pair
as ai = (p1 , . . . , pm ) where m is the number of permissions requested by each
sample app-pair. Each pj = pjq ∨ pjr where pjq and pjr are the permissions
requested by the corresponding apps aiq and air in the app-pair ai . The final
classifier assumes naive Bayes assumption for the feature set ai . The features or
permissions used in both the feature set xi and ai are the same. For each sample
app-pair ai , the final classifier calculates:

vl = θjl (2)
j
where θjl corresponds to feature probabilities conditioned on class cl : θjl =

P (pj |cl ) = P (xj |cl ). Now v1 corresponds to likelihood of an app-pair being
non-collusive or non-malicious and v2 corresponds to likelihood of an app-pair
being collusive or malicious. In our experiment, we use a set of only colluding
app-pairs. Hence, v1 is expected to be less than v2 which is found to be so. If we
use a set of both colluding and non-colluding app-pairs then v1 > v2 implies non-
collusive or non-malicious app-pair and v1 < v2 implies collusive or malicious
app-pair respectively.
In this experiment, we train our base classifier with a training set of 5000 benign
and 3000 malicious applications. We test our base classifier on a set of 2000
benign and 207 malicious applications. We further test our base classifier on
three sets of malicious applications which consist of 1260, 247, and 154 malicious
applications obtain from [13]. We design two feature sets. The first set consists of
13 critical permissions which are requested by both Android malwares and col-
luding app-pairs. The second feature set consists of permissions used in first set
plus 13 extra permissions which are frequently requested by Android malwares.
The parameter vector is obtained as a result from training our base classifier.
The parameter vector is used by the final classifier to identify an app-pair as
malicious or non-malicious as explained in the previous section. For testing of
app-pairs, we use two sets of 120 colluding app-pairs obtain from [2].
Table 1. Permission sets A and B used as features
Permission set A Permission set B

WRITE EXTERNAL STORAGE GET TASKS
CHANGE WIFI STATE RECORD AUDIO
ACCESS WIFI STATE CHANGE WIFI STATE
READ PHONE STATE READ CONTACTS
READ HISTORY BOOKMARKS READ PHONE STATE
READ CALL LOG READ HISTORY BOOKMARKS
Figures 1 and 2 shows the TPR plot of the base classifier on the four sets of
malwares. In this experiment, set 1, set 2, set 3, and set 4 consists of 207, 1260,
247, and 154 malicious applications respectively. The base classifier is trained
with feature sets of size 13 and 26. In other words, we design two permission
sets of size 13 and 26 as feature sets for training our base classifier. True positive
rates (TPR) for all the four sets have considerably improved when we go from
feature set size 13 to 26 (Table 3). It is expected because we have included a
greater number of critical permissions. We have used only the parameter vector
obtain from training our classifier with a feature set of size 13 because colluding
app-pairs requests only these permissions. This parameter vector is used in the
identification of colluding app-pairs.
Table 2. Full permission (feature) set
Permission name
GET TASKS
CHANGE WIFI STATE
RECORD AUDIO
READ PHONE STATE
READ HISTORY BOOKMARKS
READ CONTACTS
INTERNET
WRITE EXTERNAL STORAGE
READ EXTERNAL STORAGE
ACCESS WIFI STATE
GET ACCOUNTS
ACCESS FINE LOCATION
READ CALL LOG
Table 3. TPR table for base classifier
Feature Set 1 Set 2 Set 3 Set 4

set size
13 0.58 0.73 0.66 0.59
26 0.79 0.87 0.72 0.88
Figures 3, 4, 5 and 6 show the TPR plots for our final classifier on the two
sets of colluding app-pairs. We have used two permission sets A and B as feature
sets. Figures 3 and 4 shows the effects of removing permission set A on the TPR
plot of the final classifier. We train the final classifier first with permission set
A only and then with permission set obtain by removing set A from the total
feature set. The TPR considerably increases for both the sets when we use set
A. Same is the case for permission set B as shown in Figs. 5 and 6.
Figures 7 and 8 show the TPR plots for our final classifier when we remove
two important critical permissions GET TASKS and INTERNET. As we can
see, the TPR decreases considerably when we remove these permissions. Tables 1
and 2 shows the list of permissions in the set A, set B, and the full feature or
permission set. Figure 9 shows the TPR plot when we use the full feature set.
We can see that the TPR decreases considerably. The reason behind this is that
some of the permissions or features are redundant or contribute negatively to
the decision. When we remove those features and train the final classifier with
permission set A or permission set B, the TPR increases considerably.
Fig. 1. Base classifier TPR plot 1 Fig. 2. Base classifier TPR plot 2
Fig. 3. Final classifier TPR plot 1 Fig. 4. Final classifier TPR plot 2
Fig. 9. Final classifier TPR plot 7
5 Conclusion
In this work, we propose a novel and simple technique for detecting collusive app-
pairs. We use a two-tier model. In the first step, we train a classifier for detecting
malicious applications using critical permissions as features. In the next step, we
use another classifier for detecting collusive app-pairs. We use the idea that a
collusive app-pair can carry out threats like the malicious applications. The final
classifier uses the parameter vector obtain from the first step for the detection.
The app-collusion detection system achieves a TPR of 90% and 87.5% on two
sets of colluding app-pairs. This can improve by including different kinds of
attributes as features. For instance, api calls can serve another set of features
which we can employ in our final classifier. Our approach is scalable and can be
used for large scale classification of app-pairs.
Acknowledgements. This work was funded by grant from the Visvesvaraya Fellow-
ship Scheme (Govt. of India).
References
1. Bartel, A., Bissyande, T., Klein, J., Traon, Y.L., Azert, S., Li, L.: IccTA: detecting
inter-component privacy leaks in android apps. In: Proceedings of 37th Interna-
tional Conference on Software Engineering, Florence, IEEE, pp. 280–291 (2015)
2. Blasco, J., Chen T.M.: Automated generation of colluding apps for experimental
research. J. Comput. Virol. Hack. Tech. 14, 127 (2018). https://doi.org/10.1007/
s11416-017-0296-4
3. Bosu, A., Liu, F., Yao, D., Wang, G.: Collusive data leak and more: large-scale
threat analysis of inter-app communications. In: Proceedings of Asia Conference
on Computer and Communications Security, pp. 71–85. ACM (2017)
4. Bugiel, S., Davi, L., Dmitrienko, A., Fischer, T., Sadeghi, A.R.: XManDroid: a
new android evolution to mitigate privilege escalation attacks. Technical report
TR -2011-04, Technische Universitat Darmstadt (2011)
5. Bishop, C.: Pattern Recognition and Machine Learning, pp. 1–758. Springer, New
York (2006)
6. Chin, E., Porter, A.F., Greenwood, K., Wagner, D.: Analyzing inter-application
communication in Android. In: Proceedings of the 9th International Conference on
Mobile systems, Applications, and Services, pp. 239–252. ACM (2011)
7. Faiz, M.F.I., Hussain, A.M., Marchang, N.: Detection of collusive app-pairs using
machine learning. In: Proceedings of 3rd International Conference on Consumer
Electronics ASIA, pp. 206–212. IEEE (2018)
8. Felt, A.P., Wang, H.J., Moshchuk, A., Hanna, S., Chin, E.: Permission re-
delegation: attacks and defences. In: Proceedings of the 20th USENIX Security
Symposium. USENIX Association (2011)
9. Kalutarage, H.K., Nguyen, H.N., Shaikh, S.A.: Towards a threat assessment frame-
work for apps collusion. Telecommun. Syst. 66, 417–430 (2017)
10. Marforio, C., Ritzdorf, H., Francillon, A., Capkun, S.: Analysis of the communica-
tion between colluding applications on modern smartphones. In: Proceedings of the
28th Annual Computer Security Applications Conference, pp. 51–60. ACM (2012)
11. Octeau, D., Jha, S., Derring, M., McDaniel, P., Bartel, A., Li, L., Klein, J.,
Traon, Y.L.: Combining static analysis with probabilistic models to enable market-
scale android inter-component analysis. In: Proceedings of 43rd Annual ACM
SIGPLAN-SIGACT Symposium on Principles of Programming Languages, pp.
469–484. ACM (2016)
12. Porter, A.F., Ha, E., Egelman, S., Haney, A., Chin, E., Wagner, D.: Android per-
missions: user attention, comprehension, and behaviour. In: Proceedings of the 8th
Symposium on Usable Privacy and Security, Article 3, pp. 1–14. ACM (2012)
13. Wang, W., Wang, X., Feng, D., Liu, J., Han, Z., Zhang, X.: Exploring permission-
induced risk in android applications for malicious application detection. IEEE
Trans. Inf. Forensics Secur. 9(11), 1869–1882 (2014). https://doi.org/10.1109/
TIFS.2014.2353996
14. Xie, Z., Zhu, S.: GroupTie: toward hidden collusion group discovery in app stores.
In: Proceedings of Conference on Security and Privacy in Wireless & Mobile Net-
works, pp. 153–164. ACM (2014)
15. Intel newsroom. https://newsroom.intel.com/news-releases/mcafee-labs-report-
reveals-new-mobile-app-collusion-threats/
Engineering Energy Efficient
InternetWorked Smart seNsors
(E3WSN-2019)
Comparison of Machine Learning
Techniques for Prediction Problems
Yoney Kirsal Ever1(B) , Kamil Dimililer2 , and Boran Sekeroglu3

1
Software Engineering Department, Near East University,
Nicosia, Cyprus, Mersin 10, Turkey
yoneykirsal.ever@neu.edu.tr
2
Electrical and Electronic Engineering, Near East University,
kamil.dimililer@neu.edu.tr
3
Information Systems Engineering, Near East University,
boran.sekeroglu@neu.edu.tr
Abstract. Last three decades, artificial intelligence have become

mandatory for human lives since the revolution of computers. As a part
of it, Machine Learning and the proposed techniques are used to solve the
real life problems instead of humans by using their extended capabilities
for classification, optimization, and as well as prediction. However, there
is a big mystery for the optimum technique for any kind of data and prob-
lem domain. In this paper, preliminary experiments are performed using
recent and challenging nine datasets and four popular machine learn-
ing techniques in order to determine optimum technique for prediction
domain. Obtained results are analyzed using three evaluation methods;
MSE value, EV and R2 scores. Obtained results showed that, increment
or degradation of instances in datasets do not affect the performance of
the techniques directly and neural network algorithms produce higher
and more steady prediction rates.
1 Introduction
Owning to the fact that considering the improvements in the artificial intelli-
gence in last two decades are affecting the multi-disciplinary fields electronically.
Within these developments many different and interesting proposals, frameworks
and models are available in the literature. Especially, back propagation neural
networks (BPNN) with prediction are highly acceptable and deployments have
been increased especially in the fields of healthcare systems, finance and banking,
agriculture, petroleum and gas and many more [1–6].
In their paper [1], authors suggested a gradient-based BPPN model for
improving optimization in stock price predictions. They used gradient as descent
in BPNN model to decide the parameter of learning rate, for training cycle in
order to find out the best value in the process of stock data training for obtain-
ing accuracy in prediction. According to their findings stock predictions and
https://doi.org/10.1007/978-3-030-15035-8_69
714 Y. Kirsal Ever et al.
investment calculations are based on time-series algorithms whereas the neural

network-based algorithms have been widely used as a time-series data predic-
tion algorithm. Therefore, in their research, a BPNN algorithm optimized using
Gradient Descent method as approach to predict stock price. It can be con-
cluded that the neural network-based algorithm has been known and widely
used as a time-series data prediction algorithm. Therefore, this research will use
Back propagation Neural Network algorithm optimized using Gradient Descent
method as approach to predict stock price. The authors tested their model by
using mean square error (MSE) for prediction result and data reality. According
to their results the smallest MSE value showed better results compared to larger
MSE value in predictions.
Additionally, as stated above, many different discipliners have started to use
BPNN prediction algorithms [1–4]. According to Navi et al. [2], a new approach
is suggested and improved for training efficiency of BPNN algorithms. Their
proposed algorithm modifies the gradient based search direction by introduc-
ing the value of gain parameter in the activation function. They suggested that
proposed algorithm is generic and could be implemented in almost all gradient
based optimization processes, because algorithm results enhanced the compu-
tational efficiency of training process. Their suggested algorithm’s robustness
and effectiveness are tested on a heart disease data set and they are shown by
comparing convergence rates and gradient descent methods respectively.
In [3], authors suggested a crop yield-irrigation water model, that is based
on a BPNN prediction algorithm with an improved genetic algorithm (GA).
It is used to improve to develop the yield-irrigation water model to predict
the corn yield. In their results, authors discussed that their model gave more
accurate prediction. Also, it is discussed that, the proposed model speeded up
the convergence of the network. Therefore, it is concluded that the proposed
model could be used to design irrigation systems for subsurface drip irrigations.
Similarly, another study is processed during typhoon attacks for forecasting
storms in coastal areas [4]. In this study author stated that a BPNN model
is used to predict the short-term typhoon surge and its deviation in order to
overcome the problem of exclusive and nonlinear relationships. The designed
models’ numerical results are compared with the traditional method of numerical
hydrodynamic models or empirical formula results [4]. It is concluded that the
compared resulted proved the short-term storm surge and surge deviation could
be efficiently predicted 1 to 6 h ahead using BPNN.
Apart from BPNN models and algorithms, data mining techniques, especially
Iterative Dichotomiser 3 (ID3) decision tree algorithms are widely used in liter-
ature for prediction [5–8]. Decision tree uses divide and conquer technique for
the basic learning strategy [8]. A decision tree is a flow chart-like structure in
which each internal node represents a “test” on an attribute where each branch
represents the outcome of the test and each leaf node represents a class label.
In various studies, many different algorithms of the decision tree (ID3, C4.5,
CART), their features, advantages, and disadvantages are discussed in details.
Comparison of Machine Learning Techniques for Prediction Problems 715
According to literature review searches in [5], data mining is the extraction

of concealed prescient data from substantial databases, where these techniques
understood future patterns and practices, in order to make proactive, learning
driven choices. It is clearly stated that data mining processes work efficiently
when a large amount of data is available (known as big data) [8]. In data min-
ing algorithms, grouping guideline is a procedure that arranges estimations of
target variables from estimations of attributes of variables [5]. The main objec-
tive of data mining is prediction. Therefore, in study [5], data mining is utilized
for prediction of potential criminal suspects by simply analysing data retrieved
from the existing crime scenes in the existing database for detecting crimes in
universities. In the proposed model, the data was pre-processed to get clean and
accurate data. The ID3 decision tree algorithm is used to analyze and train the
data.
Another similar study [6] has created a model that classifies students accord-
ing to their successes at the end of academic year, and finding meaningful
variables effecting their success. In their study, several classification algorithms
for constructing decision trees were compared and the statistical significance
(T-test) of the results was analyzed. They emphases that the proposed algo-
rithm showed the highest accuracy was chosen as the most successful algorithm
for modelling the academic success of students. The highest classification rate
of 79% was produced using the REPTree decision tree algorithm. However, the
average rate of classification was calculated for two models that gave the highest
total rate of classification, where a higher percentage is achieved using the model
relying on the algorithm J48.
Additionally, within last 10 years, usage of social media and prediction of
human behaviours have been emerged and increased dramatically. As stated in
[7], location prediction has been investigated by many researchers. However,
authors emphasised that there are few studies that explore the relationship
between human activities and location. In their research [7], they proposed a
decision tree algorithm that investigates how location can be predicted after
observing a series of human activities. They used more than 6000 contextual
datasets for the proposed algorithm which are obtained from college students.
Their results revealed that, given a number of human activities and personal
information, a decision tree classifier provides a set of useful rules through appro-
priate inferences about location prediction are possible.
The rest of the paper is as follows, Sect. 2 discusses existing machine learning
techniques, where experiments done and results comparisons are discussed in
Sect. 3. Section 4 discusses the detailed comparison of the results listed in Sect. 3
and Sect. 5 is conclusion and recommendations for future.
2 Machine Learning Techniques

This section briefly introduces machine learning (ML) algorithms considered
in this research namely Back-propagation (BP) Learning Algorithm, Radial
Basis Function Neural Network (RBFNN), Support Vector Regression (SVR)
and Decision Tree Regressor (DTR).
Please note that the first line of text that follows a heading is not indented,
whereas the first lines of all subsequent paragraphs are.
2.1 Backpropagation Learning Algorithm
Backpropagation is one of the most popular learning algorithm in neural net-

works both for classification and prediction problems. It updates weights of each
neuron using gradient descent algorithm and different kinds of stopping criteria
can be used such as error level, iteration number etc. for the convergence of
neural network.
Backpropagation learning algorithm was used in several real-life applications
in classification, prediction and optimization problems, and it is still the most con-
sidered algorithm for the comparison of machine learning techniques [9,10,21].
2.2 Radial Basis Function Neural Network
Radial Basis Function Neural Networks was proposed in 1988 [11] and basically
it uses radial basis functions as an activation function of hidden neurons. It
acts as universal approximation on sets and its’ strength areas are function
approximation, time series prediction and classification [12,21].
2.3 Support Vector Regression
Support Vector Regression is a form of Support Vector Machines to accept real

value outputs instead of binary numbers [13]. It is effectively used in prediction
of data while minimizing error by maximizing the margin of hyperplane to dis-
tinguish the features from each other [14,15]. There are different types of SVR
as linear and non-linear, and they can be used with different kernel functions
such as polynomial, Gaussian radial basis function etc.
2.4 Decision Tree Regressor

Primarily, Decision Trees are proposed for the classification of data by using a
divide-and-conquer strategy until final leaf [16]. Then, they are modified to be
used in regression models and, their simplicity and efficiency with large number
of variables and cases make them popular for prediction problems.
3 Experiments and Results

In this section, design of experiments, considered datasets, evaluation strategy
and obtained results will be presented in details.
3.1 Datasets and Design of Experiments

As it is mentioned above, back-propagation learning algorithm, radial basis func-
tion neural network, support vector regression and decision tree techniques are
considered in order to perform comparison. These techniques are selected accord-
ing to their efficiencies and frequently usages in prediction problems.
Recent and challenging datasets; Air Quality (AQ) [17], Wine Quality (WQ)
[18] Combined Cycle Power Plant (CCPP) [19] and Real Estate (RE) Valuation
datasets [20] are used in the evaluation of prediction results of mentioned tech-
niques. Air Quality dataset consists 14 attributes and 9358 recorded instances in
a year to predict absolute humidity. Wine Quality dataset consists two different
data; one for white wine with 4898 instances and 11 attributes, and the other
for red wine with 1599 instances and 11 attributes. Output is the quality of
the wines for both data. CCPP dataset consists 9568 instances collected during
6 years and dataset was shuffled five times for effective analysis of data. RE
dataset contains 7 attributes with one house prices which is the output and 414
instances. Thus, totally 9 datasets; one for AQ, two for WQ and 5 for CCPP
and one for RE are considered in this research independently to analyse and
compare considered machine learning algorithms.
Normalization of data into the range of 0 and 1 is obtained using Min-Max
Normalization which is given in Eq. 1:
xi − min(x)
zi = (1)
max(x) − min(x)
where zi is normalized data, xi is original data and max(x) and min(x) are the
maximum and minimum values of corresponding attribute.
Evaluation of machine learning techniques for prediction problems differs
from classification problems. In classification; accuracy, specificity, sensitivity
rates etc. can be used however in prediction problems these rates are not useful
because they are based on the correctly classified features and their correspond-
ing classes which are formed by balanced data. In prediction problems, each fea-
ture has its own real-numbered class and it is impossible the measure the success
rate of machine learning algorithm with these rates. For this reason, statistical
measures such as Mean Square Error (MSE), R2 Score and Explained Variance
(EV) are frequently used in the evaluation of machine learning algorithms [9,10]
which are directly related with features and corresponding outputs.
Mean Square Error calculates the squares of error of estimator and it is
defined as:
n
1
M SE = (Yi − Ŷi )2 (2)
n i=1
where n is the total number of samples and Yi and Ŷi are the predicted and
expected outputs of estimator.
R2 Score is variance of predictable sample from the independent sample. It
is defined as:
EVs
R2 = (3)
U Vs
where EV is defined in Eq. 2 and UV is unexplained variations of samples.

Explained Variance Score is another evaluation criteria of an estimator and
also known as the regression sum of squares. It is defined as:

EVs = (fi − ŷ)2 (4)
i=1
where fi is the predicted values and ŷ is real sample.

Training are performed using 70% of all instances of considered datasets and
remaining instances are used for testing of untrained data in order to obtain
prediction rates of machine learning techniques.
After several experiments, five-layered (1-3-1) architecture is used with Sig-
moid activation function for BP and convergence criteria set to 500 epochs.
Radial Basis Function Kernel is used in SVR and γ and values are used as
0.005 and 0.01 respectively. Learning rate and maximum epoch number set to
0.09 and 4000 respectively in RBFNN. In Decision Tree Regressor, Mean Square
Error is used as a criterion.
3.2 Results
As it is mentioned above, in this research, 8 datasets are considered to analyse
and compare 4 machine learning algorithms. Table 1 shows obtained results for
all datasets and machine learning techniques.
In Air Quality Dataset, Decision Tree Regressor and BP produced optimum
results when we consider all evaluation methods. SVR produced close results to
DTR and BP in R2 and EV Scores but MSE result is relatively higher than the
others. RBFNN produced inefficient results for all evaluation metrics. Figure 1
shows prediction graph of DTR for Air Quality Dataset.
Fig. 1. Optimum prediction by DTR for air quality dataset
In Wine Quality Dataset, DTR was unable to predict features correctly and
marked as not applicable both for Red and White group. Similarly, BP and SVR
produced close but insufficient results for each other, and optimum results are
obtained by RBFNN with a big variance. Figure 2 shows minimum prediction of
WQ dataset White group by BP.
Fig. 2. Optimum prediction by BP for white wine quality dataset
In CCPP dataset, closer results are obtained for each machine learning tech-
nique in all shuffled datasets. BP produced optimum results for all in MSE, R2
Score and EV Score. It is followed by SVR and DTR in all shuffled datasets and
RBFNN produced insufficient prediction rates. Figure 3 shows the prediction
graph of backpropagation neural network for CCPP - 2.
Fig. 3. Optimum prediction by BP for CCPP - 2
In RE dataset, SVR produced optimum result when MSE value is considered

and followed by BP with closer results. However, RBFNN produced optimum
results when EV and R2 Scores are considered. Figure 4 shows the prediction
graph of SVR that obtained minimum MSE in RE dataset.
Fig. 4. Prediction graph of SVR for RE dataset
4 Discussions
As it can be shown in Table 1, different machine learning algorithms pro-

duced different results for each dataset. For AQ dataset DTR, for WQ datasets
RBFNN, for all CCPP datasets BP and for RE dataset RBFNN produced opti-
mum results.
SVR was insufficient to produce optimum results in EV and R2 scores for
any dataset but it is observed that even BP produced higher rates, results of
SVR were close to BP. This makes SVR not the optimum but one of the most
steady algorithms considered in this research.
DTR was unable to produce prediction results for WQ datasets even though
it produced successful results for higher number of instances and features in
other datasets. It performed relatively similar results to SVR and BP but it can
produce unsteady and inconsistent results.
RBFNN produced optimum results in WQ datasets for all evaluation metrics
and RE dataset for EV and R2 scores. However, for 5 CCPP datasets, it produced
lower rates than other considered techniques for all metrics.
Highest results for AQ and 5 CCPP datasets were obtained by BP and it
also performed steady and consistent performance even not producing optimum
results in other datasets. It is obvious that the characteristics of datasets, number
of instances and attributes, the parameters of considered techniques may affect
the rates but it is observed that, neural networks are superior for the prediction
of different kind of data and they can produce more successful performance than
others.
Table 1. Experimental results
BP SVR RBFNN DTR
Dataset MSE R2 EV MSE R2 EV MSE R2 EV MSE R2 EV

−5
AQ 3.6x10−7 0.9999 1.00 2.39x10 0.9993 0.9999 9.2x10−3 0.8886 0.915 1.7 × 0.9999 1.00
10−7
WQ - 0.0146 0.3491 0.352 0.0163 0.2594 0.260 0.009 0.8921 0.909 NA NA NA
Wht
WQ - 0.0162 0.310 0.311 0.0160 0.3184 0.321 7.2 × 0.910 0.910 NA NA NA
Red 10−3
CCPP -1 2.5 × 0.9504 0.9500 3.1x10−3 0.9382 0.9380 9.3x10−3 0.887 0.916 3.6x10−3 0.9300 0.9300
10−3
CCPP -2 2.5 × 0.883 0.910 3.1x10−3 0.9390 0.9390 9.7x10−3 0.883 0.910 3.8x10−3 0.9260 0.9260
10−3
CCPP -3 2.6 × 0.9491 0.9490 3.1x10−3 0.9384 0.9380 8.1x10−3 0.9023 0.9160 3.8x10−3 0.9254 0.926
10−3
CCPP -4 2.9 × 0.9415 0.942 3.4x10−3 0.9302 0.9300 0.01 0.875 0.911 3.3x10−3 0.9340 0.9340
10−3
CCPP -5 2.7 × 0.9467 0.947 3.2x10−3 0.9367 0.9360 0.01 0.8783 0.913 3.4x10−3 0.9334 0.9330
10−3
RE 5.6x10−3 0.5976 0.6410 5.3 × 0.6180 0.6180 7.7x10−3 0.8999 0.906 8.6x10−3 0.5480 0.5490
10−3
5 Conclusions
Machine Learning Techniques are proposed to solve all kind of problems but each
technique may produce higher rates in different problem domains and datasets.
Several researches were conducted and concluded different results for this.
In this paper, preliminary experiments were performed using nine datasets
for prediction in different kind of problems and four frequently used machine
learning techniques are considered for the comparison. Obtained results were
analyzed using three evaluation methods as MSE value, EV and R2 scores.
Obtained results showed that, incrementation or degradation of instances in
datasets do not affect the performance of the techniques directly but the charac-
teristics of datasets can affect the rates directly. When the analyses of evaluation
methods performed for the optimum results obtained, it can be concluded that
neural network algorithms, which are back-propagation and radial basis func-
tion neural network in this research, produced higher rates and steady results
than decision tree regressor. Similarly, support vector regression produced steady
results but could not produce optimum rates for datasets except mean square
error for real estate evaluation dataset.
Future work will include considering new and more datasets with higher
number of instances and attributes and performing comparison by adding other
techniques beside these, such as linear regression, long-short term memory etc.
Also the training and testing ratios will be investigating by degrading training
ratio to 60%.
References
1. Utomo, D., Pujiono, Soeleman, M.A.: Stock price prediction using back propa-
gation neural network based on gradient descent with momentum and adaptive
learning rate. J. Internet Bank. Commer. 22, 1–16 (2017)
2. Nawi, N.M., Ghazali, R., Salleh, M.N.M.: The development of improved back-
propagation neural networks algorithm for predicting patients with heart disease.
In: Proceedings of Information Computing and Applications - First International
Conference, ICICA 2010, pp. 317–324 (2010)
3. Gu, J., Yin, G., Huang, P., Guo, J., Chen, L.: An improved back propagation
neural network prediction model for subsurface drip irrigation system. Comput.
Electr. Eng. 60, 58–65 (2017)
4. Lee, T.-L.: Back-propagation neural network for the prediction of the short-term
storm surge in Taichung harbor, Taiwan. Eng. Appl. Artif. Intell. 21(1), 63–72
(2008)
5. Ogunde, A.O., Ogunleye, G.O., Oreoluwa, O.: A decision tree algorithm based
system for predicting crime in the university. Mach. Learn. Res. 2(1), 26–34 (2017)
6. Mesaric, J., Šebalj, D.: Decision trees for predicting the academic success of stu-
dents. Croat. Oper. Res. Rev. 7, 367–388 (2016)
7. Lee, J.S., Lee, E.S.: Exploring the usefulness of a decision tree in predicting people’s
locations. Procedia - Soc. Behav. Sci. 140, 447–451 (2014)
8. Gupta, B., Arora, A., Rawat, A., Jain, A., Dhami, N.: Analysis of various decision
tree algorithms for classification in data mining. Int. J. Comput. Appl. 163(8),
15–19 (2017)
9. Tuncal, K., Sekeroglu, B., Ozkan, C.: Lung cancer incidence prediction using
machine learning algorithms. Accepted to be published in International Journal
of Machine Learning and Computing (2019). Student Performance Prediction and
Classification Using Machine Learning Algorithms
10. Sekeroglu, B., Dimililer, K., Tuncal, K.: Student performance prediction and clas-
sification using machine learning algorithms. In: Accepted to be published in Pro-
ceedings of 8th International Conference on Educational and Information Technol-
ogy, ICEIT 2019 (2019)
11. Broomhead, D.S., Lowe, D.: Radial basis functions, mutli-variable functional inter-
polation and adaptive networks. Complex Syst. 2, 321–355 (1988)
12. Rout M., Majhi, B., Mohapatra, U.M., Mahapatra, R.: Stock indices prediction
using radial basis function neural network. In: Proceedings of International Con-
ference on Swarm, Evolutionary, and Memetic Computing, pp. 285–293 (2012)
13. Smola, A.J., Scholkopf, B.: A tutorial on support vector regression. Stat. Comput.
14(3), 199–222 (2004)
14. Henrique, B.M., Sobreiro, V.A., Kimura, H.: Stock price prediction using support
vector regression on daily and up to the minute prices. J. Financ. Data Sci. 4(3),
183–201 (2018)
15. Azeez, O.S., Pradhan, B., Shafri, H.Z.M.: Vehicular CO emission prediction using
support vector regression model and GIS. Sustainability 10(10), 3434 (2018)
16. Dougherty, G.: Pattern Recognition and Classification, 1st edn. Springer, New York
(2013)
17. De Vito, S., Massera, E., Piga, M., Martinotto, L., Di Francia, G.: On field calibra-
tion of an electronic nose for benzene estimation in an urban pollution monitoring
scenario. Sens. Actuators B: Chem. 129(2), 750–757 (2008)
18. Zhong, P., Fukushima, M.: Regularized nonsmooth Newton method for multi-class
support vector machines. Optim. Methods Softw. 22(1), 225–236 (2007)
19. Tüfekci, P.: Prediction of full load electrical power output of a base load operated
combined cycle power plant using machine learning methods. Int. J. Electr. Power
Energy Syst. 60, 126–140 (2014)
20. Yeh, I.C., Hsu, T.K.: Building real estate valuation models with comparative app-
roach through case-based reasoning. Appl. Soft Comput. 65, 260–271 (2018)
21. Ogidan E.T., Dimililer K., Ever Y.K.: Machine learning for expert systems in
data analysis. In: 2nd International Symposium on Multidisciplinary Studies and
Innovative Technologies, ISMSIT 2018 (2018)
Tailoring Micro-solar Systems
to Heterogeneous Wireless Sensor
Networks
Stefano Calabrò, Roberto Gagliardi, Fausto Marcantoni, Matteo Micheletti,

Alessandro Pacini, and Andrea Piermarteri(B)
University of Camerino, Camerino, Italy

{stefano.calabro,roberto.gagliardi,fausto.marcantoni,
matteo.micheletti,andrea.piermarteri}@unicam.it,
alessandro.pacini@studenti.unicam.it
Abstract. Energetic needs of wireless sensor networks (WSNs) have

been thoroughly studied. Among the most important results, cluster-
ing protocols are able to reduce significantly energy consumption in
these networks. In the last few years though, focus has also been put on
energy harvesting for WSNs. With energy harvesting researchers aim to
reach energy neutrality, which means the network only runs on harvested
energy. Many papers propose design options for energy harvested WSN,
but they only focus on ad-hoc solutions, homogeneous WSNs, or pose
other limitations. In this paper we propose a new approach. We study
the energetic need of a heterogeneous WSN clustered with a known algo-
rithm (REECHD) through simulation, in order to calculate the minimum
and ideal energy to harvest for a given network. Given that, we design
an appropriate micro-solar power system to achieve energy neutrality.
1 Introduction
Wireless sensor networks (WSNs) are a fundamental part of what we call Inter-
net of Things. Through WSNs we are able to generate data we have never been
able to create before, both in terms of quality and quantity. Unfortunately, to
achieve the most from this technology we need these networks to be deployed
wherever the physical phenomenon to monitor happens, and we often need them
to run for long periods. These factors lead researchers to consider energy one
of the key constraints in WSNs [11]. Clustering algorithms have been intro-
duced in WSNs to address this challenge. This approach organises the devices
into sets (clusters). Each of them has a cluster head that gathers data from its
nodes and communicates with other cluster heads in order to report data to
a centralised data sink [10]. While clustering optimises nodes energy consump-
tion, engineers introduced energy harvesting to provide wireless sensor nodes
more energy. Micro-solar power systems are one of the most used harvesting
techniques [8], due to the need to equip nodes with very compact components.
Thanks to micro-solar power systems nodes can recharge their batteries using
https://doi.org/10.1007/978-3-030-15035-8_70
Tailoring Micro-solar Systems to Heterogeneous Wireless Sensor Networks 725
solar power, thus increasing network lifetime. Many efforts have been put in this
field in order to get to the point where the network is able to perform its task
with harvesting as its only energy supply: this result is called energy neutrality
[3,4]. As we stress in Sect. 2, researchers achieved energy neutrality through ad-
hoc designs. This poses constraints and limitation to WSN developers, in terms
of hardware, software, number of nodes and network structure. To mitigate these
limitations we propose a new approach to energy harvesting WSN design. We
decide to base our work on widely-known and accepted clustering protocols for
WSNs. This way, we achieve the best results in terms of energy consumption
while maintaining a general purpose network architecture. Then, we study the
energy supply the network needs to reach energy neutrality and we derive a
suitable micro-solar power system. In Sect. 3 we define four different network
configurations and we simulate REECHD protocol on them to measure their
energy needs. Then, in Sect. 4 we use such information to design a solar harvest-
ing system with the minimal dimensions of batteries and solar panels needed
to achieve energy neutrality. In Sect. 5, we show how much energy the network
needs to harvest to reach energy neutrality in the four networks aforementioned,
as well as the necessary hardware. Finally, in Sect. 6 we show the implications
of the result we achieved, and we propose some future works.
2 Related Works
In this section we analyse the current state of the art in terms of clustering
algorithms, energy harvesting techniques, and the combination of these 2 features
in WSNs.
2.1 Classic WSN Clustering Algorithms

Clustering is one of the most widely used techniques to improve the performance
of WSNs. Clustering protocols aim to group network nodes into sets and select
one node in each set (i.e. cluster) to be the cluster head (CH). The tasks of a
CH are:
• receiving data sensed from the nodes of its cluster;
• optionally aggregate the received data;
• forward the data towards one or more base stations (BS), a node with high
storage capabilities inside the network.
In the following, we describe some well-known clustering protocols for homoge-
neous and heterogeneous WSNs. A WSN is referred to homogeneous if all the
nodes in it have the same capabilities (i.e. battery, transmission rate...). This is
not the case in heterogeneous WSN, where nodes have different capabilities.
One of the first clustering protocols to appear in the literature is Low Energy
Adaptive Clustering Hierarchy (LEACH) [7]. LEACH does not use the node
residual energy to elect CHs. Instead, LEACH cluster election phase is based
on a randomised algorithm which aims to evenly distribute the energy load
726 S. Calabrò et al.
between neighbour sensor nodes. Moreover, if a node gets the CH role, it can’t
be re-elected as CH in the next cluster election. The communication between
CHs and the BS is single-hop.
HEED [17] is another well-known clustering protocol which aim to produce
clusters with the same size. HEED uses node residual energy to perform CH
election in the election phase. Then, neighbour nodes become member nodes
by joining the least costly (usually the closest) CH. In the steady phase each
member node sends a message to its CH. Then, CHs forward the messages to
the BS. It is worth to mention that HEED algorithm prevents two nodes within
the same transmission radius to simultaneously become CHs. HEED introduces
the notion of “round” as a grouping of subsequent TDMAs. A TDMA is the
time between a node starting to collect data and the data reaching the BS.
A problem which affects HEED is the cost of the CHs election phase, which
can decrease the lifetime of the nodes by depleting their energy too fast. ER-
HEED [15] aims to improve the performance of HEED by rotating the CHs role
inside clusters. Clusters are initially formed through a classical HEED election
and formation phase, while in the next rounds each CH choose the next CH as the
cluster member node with the highest energy left. HEED cluster head election
has to be repeated only if a sensor node inside a cluster depletes its battery
completely. ER-HEED rotation algorithm makes it more effective in terms of
network lifetime of HEED.
Rotated Unequal HEED (RUHEED) [1] produces unequal-sized clusters by
decreasing the competition radius of nodes near to the BS, thus reducing the
hot-spot problem. Moreover, it adds a rotation phase after cluster election and
formation. Rotation occurs when the current CH designates the next CH among
its cluster members basing on their residual energy. The node with the highest
residual energy becomes the next CH. Rotation avoids re-election of the CHs,
reducing the communication overhead and thus postponing the death of the net-
work. Cluster election and cluster formation take place when any of the sensors
dies.
DEEC (distributed energy-efficient clustering algorithm for heterogeneous
WSNs) [12] is a clustering protocol which takes into account the energy het-
erogeneity of the network. In fact, the probability of each node to become CH
is based on the ratio between the residual node energy and the whole network
average energy. Thus, nodes with both high residual and initial energies have
more chances to become CHs.
REECHD [10] is another clustering algorithm for heterogeneous WSNs. A
peculiarity of REECHD is that it considers the transmission rate of the nodes in
the CH leader election. In particular, the election process considers the average
transmission rate of the node to estimate the extent of work it induces on its
neighbours. In fact, nodes with higher transmission rates induce less work on
neighbours and thus they should have a higher probability of becoming CH. On
the other hand, nodes with low transmission rates should have less probability of
becoming CH. That is because they generate little intra-traffic communication
inside their cluster when they are not CH.
2.2 WSN Energy Harvesting Techniques
Energy harvesting in WSNs is performed through a number of different sources.

These include solar, wind, thermal, vibration, and more ([2]).
In our study we decide to focus on solar energy because it is the most popular
harvesting source ([9,14,18], and many others), due to its efficiency, relatively
low costs and ease of use.
For example, the authors of [14] present a detailed case study of a harvested
energy WSN using solar energy to achieve energy independence. They provide a
systematic approach to building micro-solar power subsystems for WSN nodes.
While doing this, they also stress the challenges involved in the development of
harvested energy WSN powered by solar energy, as panel dimensions, battery
capacity, geographic position, time of the day and time of the year.
2.3 Clustering Algorithms for Energy Harvesting WSNs
MENC (Multi-hop Energy Neutral Clustering [16]) is a multi-hop clustering pro-

tocol for energy harvesting WSNs, with the goal of perpetual network operation.
To achieve that, MENC analyses the energy consumption of intra-cluster and
inter-cluster communication to find the best energy neutrality constraints. By
respecting such constraints, every node can work in an energy neutral state, thus
achieving perpetual network operation. The protocol mathematically derives
the minimum network data transmission cycle thanks to convex optimisation
techniques in a scenario of maximal network information gathering. In addi-
tion of ensuring perpetual network operation, MENC also improves the network
throughput with respect to traditional clustering protocols for non-harvesting
WSNs.
In [19] the authors study the relation between WSN clustering algorithms and
the presence of energy harvesting nodes. More specifically, assuming a clustered
network and a certain number of energy harvesting nodes, they manage to obtain
the optimal position for cluster heads and energy harvesting nodes.
These and many other works focus on the development of ad-hoc clustering
algorithms for harvested WSNs. This ends in a lack of generality in the field.
Also, in [3] authors analytically calculate the minimum energy harvesting
requirement for nodes in a generic clustered WSN, but they limit their study to
homogeneous WSNs.
In our study, instead, we assume a fixed clustering algorithm on a generic
heterogeneous WSN. Given that all the nodes are energy harvesting nodes, we
derive which kind of batteries and solar panel a node needs to guarantee energy
neutrality for that given network.
3 Scenario
We want to measure the energy we need to harvest to reach energy neutrality
on four different network configuration. To do this, we use the REECHD [10]
algorithm to cluster each network then for every node we get the maximum
and average energy consumption over all the Rounds until first node dies. The
energy depleted is the energy needed by each node for communication, network
synchronisation, and data collection, aggregation and forwarding to the BS.
We choose REECHD because it is one of the most optimized algorithms, and
because it is based on the well known MIT radio model [7]. The MIT model cal-
culates the transmission and reception energy consumed in sending and receiving
a data packet of k bits over a distance d as follows:
ET x (k, d) = k(Eelec + Ea dn ) (1)
ERx (k) = k(Eelec ) (2)
where Eelec = 50 nJ/bit is the energy consumption of a sensor transceiver cir-
cuit and Ea is the sensor amplification energy which depends on the distance
between a sender node and a receiver. When d < d0 = 75 m, Ea becomes Ef s =
10 pJ/bit/m2 while when d ≥ d0 = 75 m Ea becomes Emf = 0.0013 pJ/bit/m4 .
First, we define four different networks where nodes are deployed on a fixed
size square area. Nodes are not mobile i.e. they are stationary, while the BS is
located at the centre of the area.
• hW SN1 Area 100 m × 100 m, 100 nodes;
• hW SN4 Area 1000 m × 1000 m, 100 nodes.
We vary the Area size and keep the number of nodes fixed, to lower the density
of the network and increase the average distance between the nodes. Moreover,
we consider two different scenarios for each area: one in which the data sensed
from the sensors is not aggregated when sent to other nodes toward the BS, and
another where such data is aggregated. An aggregation value equal to 0 means
that all messages are sent to the BS, whereas an aggregation value of 1 results
in only one message sent to the BS (due to data compression).
In hW SN1 nodes are uniformly distributed inside the network, while in
hW SN2 , hW SN3 and hW SN4 the same nodes are deployed after different lin-
ear changes of coordinates. Due to the large size of the networks, instead of
implementing the REECHD protocol on top of hardware sensors we use a Java
based simulation tool specifically designed to run clustering protocols in hWSNs
networks. The parameters used to simulate the four scenarios described at the
top of this section are listed in Table 1.
After the simulation, we have all the data we need to define the minimum
hardware requirements as shown in the next section. Table 2 shows the energy
depleted by the nodes of network hW SN1 with no aggregation and nodes com-
petition radius set to 25 m. Emax is the maximum energy while Eaverage is the
average energy depleted by that node for each Round.
Table 1. Simulation parameters
Simulation parameters
Parameters Values
Network grid 100 × 100, 200 × 200
500 × 500, 1000 × 1000
BS (50, 50), (100, 100)
(250, 250), (500, 500)
Eelec 50 nJ/bit
Ef s 10 pJ/bit/m2
Emp 0.0013 pJ/bit/m4
Homogeneous Tx rate 2000
Homogeneous energy 1J
Heterogeneous Tx rate (500...2000)
Heterogeneous energy (1...4 J)
R0 25 m, 50 m, 125 m, 250 m
Number of nodes 100
Aggregation rate 0,1
Table 2. Extract of the energy depleted by the nodes of a hWSN
ID Coordinates (m, m) Emax (J) Rate Emax (J) Eaverage (J)

n1 (12.380, 39.037) 1 2000 2.11 ∗ 10−3 1.55 ∗ 10−3
n2 (8.780, 57.928) 1.596 1626 2.92 ∗ 10−2 3.3 ∗ 10−3
n3 (44.465, 51.654) 1 2000 4.58 ∗ 10−2 2.89 ∗ 10−3
n4 (68.603, 92.540) 2.232 1965 1.52 ∗ 10−2 2.41 ∗ 10−3
... ... ... ... ... ...
4 Energy Harvesting Model

In this section, we give the details about the harvesting model we define starting
from the metrics we get from the simulation of our four scenarios. We choose
photo-voltaic, i.e. solar panels, as the source of our harvesting.
To perform a real-life study of our scenario, we first need to define the actual
time duration of a simulation round. We fix such time to 30 s, thus having
2880 samples per day for each sensor. Knowing this, we can compute a node
daily energy depletion En daily as En average multiplied by 2880. We choose
to consider the nodes average consumption. This is enough for our case since
we assume no more than a full week without solar irradiation. Each sensor has
to be equipped with an output regulator, an electric energy storage (i.e. the
rechargeable UBP053048/PCM Ultralife Batteries of Fig. 1), an input regulator
and a solar panel with a slope angle of 0◦ . The energy which is needed to store
into each battery is computed as the product of En daily and 7 divided by the
output regulator efficiency (in our case, 50%), divided by the maximum dis-
charge percentage for the batteries (fixed to 80% of initial capacity) as shown in
Fig. 1. Finally, we can compute the minimum amount of energy the solar panel
have to output in order to achieve the neutral energy consumption goal. This is
the ratio between the energy to store into each battery and the input regulator
efficiency (fixed to 70%) divided by the batteries conversion efficiency which is
equal to 99%.
Fig. 1. UBP053048/PCM Ultralife battery specifications.
5 Results
In this section we show the results of our study. To get our results, we start from
existing climatic historical data available on the internet1 . We need this data to
estimate the average number of sunny days over the months of one year.
It is worth mentioning that we measure the minimum hardware require-
ments starting from the maximum energy needs we get from the simulations
(see Sect. 3). In fact, not all the nodes have the same energy needs. Nodes near
the BS, for instance, usually deplete their energy faster because they have to
forward data from the lower levels of the network. So, considering the highest
energy needs we get from each simulation, we can ensure that every node receives
a sufficient energy input from its harvester (i.e. the solar panel).
By looking at Table 3, we observe that both 100 × 100 and 200 × 200 scenarios
have similar energy needs in order to reach neutral energy consumption. In partic-
ular, they need a 10 cm × 10 cm panel with a peak power of 1 Wp both when the
aggregation is equal to 1 (A) and when it is equal to 0 (NA). Differently, if we con-
sider the 500 × 500 scenario we can appreciate a big difference between the A and
NA cases. In fact, the former is way less energy-hungry of the latter, which needs
a three times bigger solar panel and eight times more batteries to achieve neutral
energy consumption. Finally, in the 1000 × 1000 case, we can see that we need a
1
http://re.jrc.ec.europa.eu/pvg tools/en/tools.html.
solar panel of 1,16 m2 area and a 24 batteries pack for each sensor. This makes
clear that the long distances between the nodes affects the energy need in dispro-
portionate way, raising the minimum energy harvesting requirements to 31,78 Wh.
It is clear that the amount of harvested energy is much more than the actual
need of the nodes. One of the reasons is that there are times of the day with no
or few energy output from the harvester (i.e. at night). The graphs in Figs. 2, 3,
4 and 5 show the gap between the energy output and the energy not captured.
Table 3. Solar panels area needed for each node, according to the network config-
uration. A denotes an aggregation rate of 1, NA denotes an aggregation rate of 0.
Network Power (Wp) Area (cm2 ) Batteries Energy

hW SN1 NA 1 100 1 2,8
hW SN1 A 1 100 1 2,8
hW SN2 NA 1 100 1 2,8
hW SN2 A 1 100 1 2,8
hW SN3 NA 9 900 8 22,4
hW SN3 A 3 300 1 2,8
hW SN4 NA 116 11600 24 67,2
hW SN4 A 35 3500 11 30,8
Fig. 2. 100 × 100 with an aggregation Fig. 3. 200 × 200 with an aggregation
rate of 0. rate of 0.
Fig. 4. 500 × 500 with an aggregation Fig. 5. 1000 × 1000 with an aggregation
rate of 0. rate of 0.
6 Conclusions
In this work, we studied the energetic needs of four different hWSNs clustered
with the same algorithm (REECHD) in order to compute the amount of energy
to harvest so that they could reach an energy neutrality state. To do this, we
found the actual energy consumed from each node during the network life cycle
thus we used such information to determine the minimum amount of energy
to harvest for reaching our goal. As a future work, we propose to extend our
approach to different energy harvesting sources (piezoelectric, wind, ... ), and
to measure performances on testbeds with real hardware and advanced software
frameworks as [5,13] or [6]. Moreover, we want to collect enough data from other
scenarios as a base to develop a solid analytical model able to calculate hWSNs
energy harvesting needs for any given network.
References
1. Aierken, N., Gagliardi, R., Mostarda, L., Ullah, Z.: Ruheed-rotated unequal clus-
tering algorithm for wireless sensor networks. In: 29th IEEE International Con-
ference on Advanced Information Networking and Applications Workshops, AINA
2015 Workshops, Gwangju, South Korea, 24–27 March 2015, pp. 170–174 (2015)
2. Anisi, M.H., Abdul-Salaam, G., Idris, M.Y.I., Wahab, A.W.A., Ahmedy, I.: Energy
harvesting and battery power based routing in wireless sensor networks. Wireless
Netw. 23(1), 249–266 (2017)
3. Besbes, H., Smart, G., Buranapanichkit, D., Kloukinas, C., Andreopoulos, Y.: Ana-
lytic conditions for energy neutrality in uniformly-formed wireless sensor networks.
IEEE Trans. Wireless Commun. 12(10), 4916–4931 (2013)
4. Chen, W., Andreopoulos, Y., Wassell, I.J., Rodrigues, M.R.D.: Towards energy
neutrality in energy harvesting wireless sensor networks: a case for distributed
compressive sensing? In: 2013 IEEE Global Communications Conference (GLOBE-
COM), pp. 474–479, December 2013
5. Dulay, N., Micheletti, M., Mostarda, L., Piermarteri, A.: PICO-MP: de-centralised
macro-programming for wireless sensor and actuator networks. In: 2018 IEEE 32nd
(AINA), pp. 289–296, May 2018
6. Gummadi, R., Gnawali, O., Govindan, R.: Macro-programming wireless sensor
networks using Kairos. In: Proceedings of First IEEE International Conference on
Distributed Computing in Sensor Systems, DCOSS 2005, Marina del Rey, CA,
USA, 30 June–1 July 2005, pp. 126–140 (2005)
7. Heinzelman, W.R., Chandrakasan, A., Balakrishnan, H.: Energy-efficient commu-
nication protocol for wireless microsensor networks. In: Proceedings of the 33rd
Hawaii International Conference on System Sciences, vol. 8, HICSS 2000, p. 8020,
Washington, DC, USA. IEEE Computer Society (2000)
8. Jeong, J., Jiang, X., Culler, D.: Design and analysis of micro-solar power systems
for wireless sensor networks. In: 2008 5th International Conference on Networked
Sensing Systems, pp. 181–188, June 2008
9. Li, Y., Shi, R.: An intelligent solar energy-harvesting system for wireless sensor
networks. EURASIP J. Wirel. Commun. Netw. 2015(1), 179 (2015)
10. Micheletti, M., Mostarda, L., Piermarteri, A.: Rotating energy efficient clustering
for heterogeneous devices (REECHD). In: Barolli, L., Takizawa, M., Enokido, T.,
Ogiela, M.R., Ogiela, L., Javaid, N. (eds.) 32nd IEEE International Conference on
Advanced Information Networking and Applications, AINA 2018, Krakow, Poland,
16–18 May 2018, pp. 213–220. IEEE Computer Society (2018)
11. Miorandi, D., Sicari, S., Pellegrini, F.D., Chlamtac, I.: Internet of things: vision,
applications and research challenges. Ad Hoc Netw. 10(7), 1497–1516 (2012)
12. Qing, L., Zhu, Q., Wang, M.: Design of a distributed energy-efficient clustering
algorithm for heterogeneous wireless sensor networks. Comput. Commun. 29(12),
2230–2237 (2006)
13. Stanford-Clark, A., Truong, H.L.: MQTT for sensor networks (MQTTS) specifica-
tions, October 2007
14. Taneja, J., Jeong, J., Culler, D.: Design, modeling, and capacity planning for micro-
solar power sensor networks. In: 2008 International Conference on Information
Processing in Sensor Networks (IPSN 2008), pp. 407–418, April 2008
15. Ullah, Z., Mostarda, L., Gagliardi, R., Cacciagrano, D., Corradini, F.: A compari-
son of heed based clustering algorithms – introducing er-heed. In: 2016 IEEE 30th
(AINA), pp. 339–345, March 2016
16. Yang, L., Lu, Y.-Z., Zhong, Y., Wu, X.G., Yang, S.X.: A multi-hop energy neu-
tral clustering algorithm for maximizing network information gathering in energy
harvesting wireless sensor networks. In: Sensors (2015)
17. Younis, O., Fahmy, S.: Heed: a hybrid, energy-efficient, distributed clustering app-
roach for ad hoc sensor networks. IEEE Trans. Mob. Comput. 3(4), 366–379 (2004)
18. Yu, H., Yue, Q.: Indoor light energy harvesting system for energy-aware wireless
sensor node. Energy Procedia 16, 1027–1032 (2012). 2012 International Conference
on Future Energy, Environment, and Materials
19. Zhang, P., Xiao, G., Tan, H.: Clustering algorithms for maximizing the lifetime
of wireless sensor networks with energy-harvesting sensors. Comput. Netw. 57,
2689–2704 (2013)
Distributing Energy Consumption
in Multi-interface Series-Parallel
Networks
Alessandro Aloisio1,2 , Alfredo Navarra3(B) , and Leonardo Mostarda4

1
School of Advanced Studies, Gran Sasso Science Institute, L’Aquila, Italy
2
Department of Information Engineering, Computer Science and Maths,
University of L’Aquila, L’Aquila, Italy
alessandro.aloisio@univaq.it
3
Department of Mathematics and Computer Science, University of Perugia,
Perugia, Italy
alfredo.navarra@unipg.it
4
School of Science and Technology, University of Camerino, Macerata, Italy
leonardo.mostarda@unicam.it
Abstract. Nowadays, communication networks are composed by het-

erogeneous devices that can communicate by means of multiple inter-
faces. By choosing which interfaces to activate (switch-on) at each device,
several connections might be established. That is, the devices at the end-
points of each connection share at least one active interface. This is at
the basis of a new but well-investigated model referred in the litera-
ture to as Multi-Interface networks. In this paper, we consider a new
variant of the original model where each device is limited to activate
at most a fixed number p of its available interfaces. In particular, we
consider the so-called Coverage problem. Given a network G = (V, E),
nodes V represent devices, edges E represent connections that can be
established. The aim is to activate at most p interfaces at each node
in order to establish all the connections defined by E. A connection is
established whenever the two endpoints activate one common interface.
Recently, the problem has been proved to be NP-hard even for the basic
case of p = 2. Then, various resolution algorithms have been provided
for different graph topologies. Here we keep on investigating on the case
p = 2 when the underlying graph represents a so-called Series-Parallel
network. We provide an optimal resolution algorithm based on dynamic
programming for this intriguing graph class.
1 Introduction
As technology advances and hardware costs are reduced, very powerful devices
are available for a wide range of applications. Moreover, heterogeneous devices
The work has been supported in part by the European project “Geospatial based
Environment for Optimisation Systems Addressing Fire Emergencies” (GEO-SAFE),
contract no. H2020-691161, and by the Italian National Group for Scientific Compu-
tation (GNCS-INdAM).
https://doi.org/10.1007/978-3-030-15035-8_71
Multi-interfaces on Serial-Parallel Networks 735
may communicate to each other by means of different protocols and inter-

faces. The connection among heterogeneous devices might result as a funda-
mental means for the communication among different local area networks that
all together form a wide area network, or the Internet. In this paper, we are inter-
ested in networks composed by heterogeneous devices that exploit different com-
munication interfaces in order to establish desired connections. It is very common
to find devices equipped with Bluetooth, Wi-Fi and 4G interfaces but very few
applications usually take advantage of such available heterogeneity. Selecting the
best interfaces for specific connections depends on several factors. Namely, the
choice might depend on the availability of a specific interface on some devices,
the required communication bandwidth, the cost (in terms of energy consump-
tion) for maintaining an active interface, the neighborhood, and so forth. In [17],
for instance, the authors provide an experimental study on the choice of using
either Bluetooth or Wi-Fi for the transmission of data packets among various
smart-phones. The choice has been based on the energy consumed for accom-
plishing the required communications by varying on the size of the packets. Since
devices are usually portable or mobile, a lot of effort must be devoted to energy
consumption issues in order to prolong the network lifetime. In fact, the failure
of a device due to draining batteries is something that could be delayed if suit-
able solutions are provided when building the connections of a desired network
in accordance to specific requirements. This introduces challenging and natural
optimization problems that must take care of different parameters at the same
time. In general, a network of devices will be described by a graph G = (V, E),
where V represents the set of devices and E is the set of possible connections
defined according to the distance between devices and the available interfaces
that they share. Each v ∈ V is associated with a set of available interfaces W (v).
Theset of all the possible available interfaces in the network is then determined
by v∈V W (v); we denote the cardinality of this set by k. We say that a connec-
tion is established when the endpoints of the corresponding edge share at least
one active interface. If an interface α is activated at some node u, then u con-
sumes some energy c(α) for maintaining α as active, and it provides a maximum
communication bandwidth b(α) with all its neighbors that share interface α. All
the possible connections can be established by means of at least one interface.
Note that, some devices are not directly connected even though they share some
interfaces. This can be due to many factors like for instance obstacles, distances,
or used protocols. In order to provide a full instance, one should provide (if
necessary) the cost for each device to activate a specific interface, and the corre-
sponding bandwidth that can be handled. While bandwidths usually concern the
interfaces, the energy spent by each device to activate a specific interface may
vary substantially. In order to make easier the model, the costs might be referred
to the percentage of battery consumed by each device, and hence it might be
considered the same for each device with respect to a specific interface among
the whole network. Nevertheless, different assumptions may lead to completely
different problems that point out specific peculiarities of the composed networks.
736 A. Aloisio et al.
In this paper, we are interested in the so-called Coverage problem, see [19]. It
consists in finding the cheapest way to establish all the connections defined by an
input graph G, no matter the interface used to accomplish each connection, no
bandwidths requirements are provided. The problem only asks to ensure that for
each edge of G, there is a common active interface at its endpoints. The objective
is to minimize the overall cost of activation in the network. The problem has been
extensively studied even with respect to another interesting objective function
that concerns the minimization of the maximum cost on a single node, see [12].
Consistently with [4], here we consider one further constraint to the original
setting, in order to maintain under control the energy spent by single devices.
A parameter p is introduced with the meaning that no device can activate more
than p interfaces. That is, instead of finding the solution that minimizes the over-
all cost due to the activation of the interfaces along the whole network, or the
one that minimizes the maximum cost spent by a single node, the aim becomes
to minimize the overall cost subject to constraint p. This is very important for
practical purposes as it is a step towards the study of multi-interface networks in
distributed setting. Similarly to [10], this is a new approach to maintaining bal-
anced the energy consumption of the single nodes. This approach clearly helps
in prolonging the lifetime of the network. A possible context where our theoret-
ical results can be exploited may concern military applications, and specifically
the tactical networks provided with new software technologies [3,23]. In fact,
this type of networks suffers for the employing of resource-constrained mobile
devices. Clearly, one of the main resources that has to be preserved is the battery
power of each device, this implies that the suitable usage of the interfaces might
be crucial for prolonging the network lifetime.
1.1 Related Work
Multi-interface networks have been extensively studied during the last years in a
variety of contexts, usually focusing on the benefits of exploiting multiple inter-
faces of each device. Many basic problems of standard network optimization are
reconsidered in such a setting [7,10], in particular, focusing on issues related to
routing [13] and network connectivity [9,15]. Combinatorial problems on multi-
interface wireless networks have been studied starting from [8]. That paper, as well
as [4,12,19], investigate the Coverage problem and its variants described above.
Connectivity issues have been addressed in [5,22]. The problem consists in
finding the cheapest way to ensure the connectivity of the entire network. In
other words, it aims to find at each node a subset of the available interfaces that
must be activated in order to guarantee a path between every pair of nodes in
G while minimizing the overall cost of the interfaces activated among the whole
network. As for the case of Coverage, another studied objective function is that
of minimizing the maximum cost paid on a single node [12]. The Connectivity
problem corresponds to a generalization of the well-know Minimum Spanning
Tree problem. In fact, costs are not on the edges but on the interfaces hold by
the nodes. It may happen that the same interface is used by a node to establish
several connections, hence saving energy. This property of multi-interface net-

works reveals the advantage and the higher complexity of the studied problems.
In [22], the attention has been devoted to the so called Cheapest path problem.
Here, the goal is to find a minimum cost subset of available interfaces that can
be activated in some nodes in order to guarantee a connecting path between two
specified nodes of the network. This problem corresponds to a generalization of
the well-know Shortest Path problem between two nodes in standard networks.
Interestingly, this is one of the few problems studied in the multi-interface con-
text that maintains its computational complexity close to its standard version
and hence it can be optimally solved in polynomial time.
Another interesting study [21] investigates a basic problem like the Maximum
Matching. As in its classical version, the problem asks for the maximum subset of
connections that can be established concurrently without sharing any common
node. That is, a solution is provided by a set of disjoint edges of the input graph,
and hence each node appears in the solution at most once. The difficulty of the
problem comes by observing the Maximum Matching problem in the field of
multi-interface networks reveals another interesting peculiarity of this kind of
networks. An instance of the problem is simply provided by an input graph G
without specifying costs and bandwidths. The problem becomes in fact already
difficult to be solved since one must consider that when a solution includes
two edges established by means of the same interface, they cannot be directly
connected by another edge, as otherwise the connection represented by this third
edge would be established as well since both its endpoints share the same active
interface, hence making invalid the current solution.
Bandwidth constraints have been addressed in [6,11], by studying flow prob-
lems on multi-interface networks. In those papers, each interface has been asso-
ciated with one further parameter expressing the bandwidth that the interface
can manage. The Maximum Flow problem as well as the Minimum Cost Flow
problem aim to guarantee a connection between two given nodes, taking into
account bandwidth constraints. The Maximum Flow problem consists in find-
ing the maximal possible bandwidth between two selected nodes. In detail, all
the interfaces of the network are considered as active, so that all the allowed
connections are established. Then, a suitable flow function has been investigated
that guarantees the maximum communication bandwidth between the two given
nodes. This problem corresponds to a generalization of the Maximum Flow prob-
lem in standard networks. The Minimum Cost Flow problem consists in estab-
lishing a communication sub-network between two given nodes of minimum cost
in terms of energy consumption, while guaranteeing a minimum communication
bandwidth B. In other words, the problem asks for the minimum cost set of
active interfaces among the input network in such a way that a certain node s
is guaranteed to exchange data with another specified node t with bandwidth
of at least B. In general, the solution is not a path between s and t, but a
more complex graph consisting of nodes with active interfaces might be required
according to the topology and the available interfaces. This problem corresponds
to a generalization of the Minimum Cost Flow problem in standard networks.
1.2 Our Results
In this paper, we are interested in the Coverage problem with the further con-
straint that each node of the network can activate at most 2 interfaces. According
to [4], from now on the problem is denoted by CMI(2) and we consider the so-
called unbounded case where there is no bound on the number k of interfaces
available over all the network. Since in general it has been already proved in [4]
that the problem is NP-hard, it is worth investigating on specific graph topolo-
gies. Basic ones like paths, rings, trees, complete graphs and complete bipartite
graphs have been already addressed and optimally solved in [4]. Here we keep on
investigating on a wide graph class, namely the Series-Parallel graphs [14]. The
formal definition of such intriguing topology will be given in the next section.
Informally, Series-Parallel graphs are graphs with two distinguished nodes called
terminals, formed recursively by two simple composition operations. Such graphs
have been largely investigated in the last decades, mainly for their relationship
with series and parallel electric circuits. We provide a polynomial time dynamic
programming algorithm that optimally solves CMI(2) on Series-Parallel graphs.
1.3 Outline
The next section provides definitions and notation in order to formally describe
the CMI(p) problem.
Section 3 is the core of the paper, containing the formal definition of Series-
Parallel graphs and the resolution algorithm that works in polynomial time.
Finally, Sect. 4 contains conclusive remarks and a discussion of interesting
open problems.
2 Preliminaries
For a graph G, we denote by V its node set, by E its edge set. Unless otherwise
stated, the graph G = (V, E) representing the network is assumed to be undi-
rected, connected, and without multiple edges and loops. For each node v ∈ V ,
let deg(v) be the degree of v, and Δ be the maximum degree among all the
nodes. When considering the network graph G, we simply denote the number of
its nodes and edges by n and m, respectively. A global assignment of the inter-
faces to the nodes in V is given in terms of an appropriate interface assignment
function W , as follows.
Definition 1. A function W : V → 2{1,2,...,k} is said to cover graph G if W (u) ∩
W (v) = ∅, for each {u, v} ∈ E.
The cost of activating an interface for a node is assumed to be identical for all
nodes and given by a cost function c : {1, . . . , k} → R+ , i.e., the cost of interface
i is denoted as ci . The considered CMI(p) optimization problem is formulated
as follows.
CMI(p): Coverage in Multi-Interface Networks

Input: A graph G = (V, E), an allocation of available interfaces W : V →
2{1,...,k} covering graph G, an interface cost function c : {1, . . . , k} →
R+ , an integer p ≥ 1
Solution: If possible, an allocation of active interfaces WA : V → 2{1,...,k} covering
G such that for all v ∈ V , WA (v) ⊆ W (v) and |WA (v)| ≤ p; Otherwise, a
negative answer
Goal : Minimize the total cost of the active interfaces, c(WA ) =

v∈V i∈W (v) ci
A
Note that we can consider two variants of the above problem: The cost func-
tion c can span over R+ , or c(i) = 1, with i = 1, . . ., k (unit cost case). In both
cases we assume k ≥ 2, since the case k = 1 admits an obvious unique solution
(all nodes must activate their unique interface).
In [4] (and similarly to the approach adopted in [1]), it has been shown that
CMI(p) is a special case of the more general CMI(∞) problem (see [10,19]),
where each node cannot activate more than p interfaces. Interestingly, on the
one hand the basic variant with p = 2 results more difficult in general than
CMI(∞). On the other hand, there are special graph classes that turn out to be
much more affordable. For trees and complete graphs, for instance, CMI(∞) has
been proved to be APX -hard and not approximable within O(log k), respectively,
while CMI(2) is polynomially solvable.
3 Series Parallel Graphs

We start by formally define what is a Series-Parallel graph (SP-graph). Subse-
quently, we present our new resolution algorithm that optimally solves CMI(2)
on SP-graphs in polynomial time.
3.1 Definition of SP-Graphs

Definition 2 ([20]). A single edge (i, j) is an SP-graph with source i and sink
j (i and j are called terminals). If G1 and G2 are two SP-graphs, then also the
graph obtained by one of the next operations is an SP-graph:
1. series composition: identify the sink of G1 with the source of G2 .
2. parallel composition: identify the source of G1 with the source of G2 and the
sink of G1 with the sink of G2 .
It is common to refer to SP-graphs without multiple edges as simple SP-

graphs. Since our problem does not present multiple edges, for the sake of sim-
plicity we omit the word simple in what follows.
On the basis of the recursive Definition 2, it is possible to represent an SP-
graph into a rooted binary tree where each leaf corresponds to an edge of the
SP-graph and an interior node is labeled with an ‘S’ (series composition) or with
a ‘P’ (parallel composition). The binary tree representation describes a way to
reconstruct the SP-graph following the rules of series and parallel compositions.
Fig. 1. A Series-Parallel graph. Fig. 2. A binary decomposition tree

of the Series-Parallel graph shown in
Fig. 1.
In particular, the subgraph G(v) (induced by the edges of the subtree rooted in
the interior node v) can be constructed by a series or a parallel composition of
the two SP-graphs G(v1 ) and G(v2 ) where v1 and v2 are the children of v in the
binary tree. The authors of [24] described an O(n) algorithm that gets a graph
G as input and returns a binary tree representation of the G if and only if the
G is an SP-graph. Moreover, it is well-known that the number m of edges of an
SP-graph satisfies the equations n − 1 ≤ m ≤ 2n − 3, and a simple proof can be
found in [20].
An example of an SP-graph, and a possible binary decomposition tree are
depicted in Figs. 1 and 2, respectively.
3.2 Solving CMI(2) on SP-Graphs
This section describes a polynomial time optimal algorithm for CMI(2) using
the dynamic programming technique. The algorithm exploits the binary tree
representation of an SP-graph provided above. Then, the algorithm solves the
subproblems on the leaves and then iteratively extends these solutions until the
root where the whole problem is solved.
Definition 3. Given a binary tree T representing an SP-graph, let v be a generic

node of T inducing the SP-graph G(v) with source i and sink j. For a subset
A, with |A| ≤ 2, of the available interfaces in i and for a subset B, with |B| ≤
2, of the available interfaces in j, we denote by f (v, A, B) the minimum cost
needed to solve the CMI(2) instance restricted to G(v) that satisfies the following
conditions:
• the active interfaces in i are those in A;

• the active interfaces in j are those in B.
Following this definition, at each node v of T we calculate the values

f (v, A, B), for every pair of subsets A and B, each of at most two of the available
interfaces in the source and the sink, resp., of G(v), via a dynamic programming
algorithm that starts from the leaves and ends at the root. If there is no solution
of the described form, then we set f (v, A, B) = +∞.
If v is a leaf, clearly one of the following conditions holds:
• f (v, A, B) = c(A) + c(B) if A ∩ B = ∅;
• f (v, A, B) = +∞ if A ∩ B = ∅;
In fact, the SP-subgraph G(v) induced by the leaf v is the edge (i, j). So,
the solution that activates the interfaces A in i and B in j is feasible, with cost
c(A) + c(B), if and only if A ∩ B = ∅.
If v is an interior node of T with label S (series composition), with left child
v1 and right child v2 , we can compute easily the minimum cost f (v, A, B) by
using the optimal values already calculated for v1 and v2 . Let i and h be the
source and the sink, of G(v1 ), and let h and j be the source and the sink of
G(v2 ). Then, f (v, A, B) is equal to the optimal value of the problem:
min f (v1 , A, C) + f (v2 , C, B) − c(C)

(1)
∀C ⊆ W (h), |C| ≤ 2
In fact, to be a proper series composition, the node h is the sink in G(v1 ) and
the source in G(v2 ). So, the cost of the active interfaces in C associated to h is
counted twice. This is the reason why we subtract c(C) in the objective function.
Analogously, we can easily compute the optimal value of an interior node v
of T with label P (parallel composition) and children v1 and v2 . Let i and j
be the source and the sink, resp., of both G(v1 ) and G(v2 ), then the value of
f (v, A, B) is equal to:
f (v1 , A, B) + f (v2 , A, B) − c(A) − c(B) (2)
As nodes i and j are the source and the sink, resp., of both G(v1 ) and G(v2 ),
then the costs of the active interfaces A and B are counted twice. This is the
reason why we subtract c(A) and c(B) in the objective function.
Using this algorithm, the optimal value of CMI(2) is the minimum value
f (root, A, B), for every possible subsets A and B of the available interfaces in
the source i and the sink j of the root of T , with |A| ≤ 2 and |B| ≤ 2. At
each node labelled with S, O((k + k2 )3 ) values f (v, A, B) are computed. In
fact, we need to test on three subsets A, B, and C of interfaces containing at
most 2 interfaces each,
and all their combinations. Instead, at each node labelled
with a P, O((k + k2 )2 ) values f (v, A, B) are computed as we need to test two
subsets A and B of interfaces containing at most 2 interfaces each, and all their
combinations. Since the number of the edges in an SP-graph is at most 2n − 3,
then, the numberofthe nodes in T is O(n). Summarizing, the dynamic algorithm
runs in O(n(k + k2 )3 ) time. We can then state the following:
Theorem 1. CMI(2) is solvable on Series-Parallel graphs in O(k 6 n) time.
In this paper, we have continued our investigation on a variant of the well-
studied Coverage problem arisen in the context of multi-interface networks. The
task asks to find the cheapest way to establish all the connections defined by an
input graph by activating suitable subsets of interfaces available at the nodes
of the network. The modification to the original model concerns the further
constraint for which each node can activate at most p interfaces. In particu-
lar we focus on the case with p = 2 and considering Series-Parallel graphs as
underlying topology. The problem is known to be NP-hard in general whereas
we could devise a polynomial time algorithm that optimally solve the problem
in the considered context. The new proposed algorithm is based on dynamic
programming.
Considering that Series-Parallel graphs admit a treewidth of at most 2, it would
be interesting to keep on investigating whether the problem can be solved in graphs
with bounded treewidth. Also planar graphs would be worth to investigate as
Series-Parallel graphs contain the class of outerplanar graphs. Moreover, as the
NP-hardness is given for general graphs of maximum degree 4, and that for max-
imum degree 2 the problem is optimally solvable in polynomial time, it remains
to investigate the case of graphs with maximum degree 3. Finally, new interesting
research lines in the context of Multi-Interface networks could be that of consid-
ering small-world [16] or greedy coloring [18] approaches, or multi-objective func-
tions. For instance, similar to [2], a requirement might be to minimize the number
of types of used interfaces subject to a bounded global cost.
References
1. Aloisio, A., Arbib, C., Marinelli, F.: Cutting stock with no three parts per pattern:
work-in-process and pattern minimization. Discrete Optim. 8, 315–332 (2011)
2. Aloisio, A., Arbib, C., Marinelli, F.: On LP relaxations for the pattern minimization
problem. Networks 57, 247–253 (2011)
3. Aloisio, A., Autili, M., D’Angelo, A., Viidanoja, A., Leguay, J., Ginzler, T., Lampe,
T., Spagnolo, L., Wolthusen, S., Flizikowski, A., Sliwa, J.: TACTICS: tactical ser-
vice oriented architecture. In: 3rd International Conference in Software Engineering
for Defence Applications, pp. 1–9 (2014)
4. Aloisio, A., Navarra, A.: Balancing energy consumption for the establishment of
multi-interface networks. In: Proceedings of the 41st International Conference on
Current Trends in Theory and Practice of Computer Science, (SOFSEM). LNCS,
vol. 8939, pp. 102–114. Springer (2015)
5. Athanassopoulos, S., Caragiannis, I., Kaklamanis, C., Papaioannou, E.: Energy-
efficient communication in multi-interface wireless networks. Theory Comput. Syst.
52, 285–296 (2013)
6. Audrito, G., Bertossi, A., Navarra, A., Pinotti, C.: Maximizing the overall end-user
satisfaction of data broadcast in wireless mesh networks. J. Discrete Algorithms
45, 14–25 (2017)
7. Bahl, P., Adya, A., Padhye, J., Walman, A.: Reconsidering wireless systems with
multiple radios. SIGCOMM Comput. Commun. Rev. 34(5), 39–46 (2004)
8. Caporuscio, M., Charlet, D., Issarny, V., Navarra, A.: Energetic performance of
service-oriented multi-radio networks: issues and perspectives. In: Proceedings of
the 6th International Workshop on Software and Performance (WOSP), pp. 42–45.
ACM (2007)
9. Cavalcanti, D., Gossain, H., Agrawal, D.: Connectivity in multi-radio, multi-
channel heterogeneous ad hoc networks. In: Proceedings of the 16th International
Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC), pp.
1322–1326. IEEE (2005)
10. D’Angelo, G., Di Stefano, G., Navarra, A.: Multi-interface wireless networks: com-
plexity and algorithms. In: Ramakrishnan, S., El Emary, I.M.M. (eds.) Wireless
Sensor Networks: From Theory to Applications, pp. 119–155. CRC Press/Taylor
& Francis Group, Boca Raton (2013)
11. D’Angelo, G., Di Stefano, G., Navarra, A.: Flow problems in multi-interface net-
works. IEEE Trans. Comput. 63, 361–374 (2014)
12. D’Angelo, G., Stefano, G.D., Navarra, A.: Minimize the maximum duty in multi-
interface networks. Algorithmica 63(1–2), 274–295 (2012)
13. Draves, R., Padhye, J., Zill, B.: Routing in multi-radio, multi-hop wireless mesh
networks. In: Proceedings of the 10th International Conference on Mobile Com-
puting and Networking (MobiCom), pp. 114–128. ACM (2004)
14. Duffin, R.: Topology of series-parallel networks. J. Math. Anal. Appl. 10(2), 303–
318 (1965)
15. Faragó, A., Basagni, S.: The effect of multi-radio nodes on network connectivity—
a graph theoretic analysis. In: Proceedings of the 19th International Symposium
on Personal, Indoor and Mobile Radio Communications, (PIMRC), pp. 1–5. IEEE
(2008)
16. Flammini, M., Moscardelli, L., Navarra, A., Pérennes, S.: Asymptotically optimal
solutions for small world graphs. In: Proceedings of the 19th International Confer-
ence on Distributed Computing (DISC). Lecture Notes in Computer Science, vol.
3724, pp. 414–428. Springer (2005)
17. Friedman, R., Kogan, A., Krivolapov, Y.: On power and throughput tradeoffs
of WiFi and Bluetooth in smartphones. In: Proceedings of the 30th International
Conference on Computer Communications (INFOCOM), pp. 900–908. IEEE (2011)
18. Gavoille, C., Klasing, R., Kosowski, A., Kuszner, L., Navarra, A.: On the com-
plexity of distributed graph coloring with local minimality constraints. Networks
54(1), 12–19 (2009)
19. Klasing, R., Kosowski, A., Navarra, A.: Cost minimization in wireless networks
with a bounded and unbounded number of interfaces. Networks 53(3), 266–275
(2009)
20. Korenblit, M., Levit, V.: On algebraic expressions of series-parallel and Fibonacci
graphs. In: Calude, C.S., Dinneen, M.J., Vajnovszki, V. (eds.) Discrete Mathemat-
ics and Theoretical Computer Science, pp. 215–224. Springer, Heidelberg (2003)
21. Kosowski, A., Navarra, A., Pajak, D., Pinotti, C.: Maximum matching in multi-
interface networks. Theor. Comput. Sci. 507, 52–60 (2013)
22. Kosowski, A., Navarra, A., Pinotti, M.: Exploiting multi-interface networks: con-
nectivity and cheapest paths. Wirel. Netw. 16(4), 1063–1073 (2010)
23. Perucci, A., Autili, M., Tivoli, M., Aloisio, A., Inverardi, P.: Distributed com-
position of highly-collaborative services and sensors in tactical domains. In: Pro-
ceedings of the 7th International Conference in Software Engineering for Defence
Applications, (SEDA), LNCS. Springer (2018, to appear)
24. Valdes, J., Tarjan, R., Lawler, E.: The recognition of series parallel digraphs.
In: Proceedings of the 11th Annual ACM Symposium on Theory of Computing
(STOC), pp. 1–12. ACM (1979)
Energy Efficient Light Routing in Utility
Network
Rosario Culmone(B) and Fabio Pagnotta(B)
University of Camerino, Camerino, Italy

{rosario.culmone,fabio.pagnotta}@unicam.it
Abstract. Nowadays, automatic meter reading is an essential technol-

ogy for customers and providers. The meter communication infrastruc-
ture is frequently subject to interference and is composed of nodes with
limited memory and processing capability. These limitations make the
development of routing protocols a challenging task. We propose a new
centralised reliable protocol based on European M-Bus [EN13757] that
uses a distributed cache with a master/slave model. This reduces the
network overhead while keeping a good throughput. We have validate
our protocol by considering a network with white Gaussian noise and
binary symmetric channels. The simulation shows a reduced fault while
keeping a good throughput and low power consumption.
1 Introduction
Every nation has billions of building that have gas, water and electricity which
need to be measured periodically. Smart metering allows the reading of resource
consumption automatically. This technology reduces the need for manual read-
ing, decreases the annual management cost and provides real-time readings.
These help the users and the provider to reduce consumption and improve the
quality of service [3]. The implementation of smart metering systems is also
required by the Directive 2009/72/EC of the European Parliament in order to
favour the engagement of the consumers in the electricity and gas supply mar-
kets. The metering system components are devices and network. Devices also
called metric meters, perform the measurement of a resource. Their electronic
part is principally made up of simple microcontrollers with sensors. These devices
have limited capability in terms of computation, memory and power usage. For
this reason, the highest cost is not the device itself but the subsystem of inter-
connection and communication. Optical fiber, WiFI, wireless GSM and others
are amongst them. The network represents the critical part in the deployment
of the reading service [2].
Although smart meter is a branch of smart grids, the networks are extremely
different. The standard for smart grid [1] are not adequate for networks of smart
meters. For instance, the OSGP [Open Smart Grid Protocol] and ISO/IEC 14908
[1] standards have many features that are inadequate or not useful for smart
meter systems. These use Meter bus (a European standard) that includes the
https://doi.org/10.1007/978-3-030-15035-8_72
746 R. Culmone and F. Pagnotta
physical [5], the link [5] (EN 13757-2) and the application layers [6] (EN 13757-
3). In particular, EN 13757-4 describes an M-bus wireless solution [4]. Although
is optional, the network layer identifies only nodes inside the network and does
not standardise any routing algorithm. Such protocol does not restrain the net-
work topology. Only the ring topology should be avoided. M-BUS is based on
a master/slave model where the supplier has the role of master and asks peri-
odically to all nodes the resource consumption. Slaves receive the request and
provide back the measured value. In most of the case, master and slave are not
directly connected but their communication is achieved via a router or repeater.
The maximum distance between a slave and a repeater is 350 m, and the max-
imum number of slaves can reach 250 [5]. This introduces the need of routing
amongst nodes. The meter communication infrastructure is frequently subject
to interference and is composed of nodes with limited memory and processing
capability. These limitations make the development of routing protocols a chal-
lenging task. Various routing approaches have been proposed in literature. In
[10,11] the authors propose an energy aware routing that is implemented as an
extension of the EN13757 Wireless M-Bus Q-mode. A cost efficient integrated
energy harvesting system powered by the available water flow was developed,
to enable operation independently of the mains grid, and eliminate the need for
battery replacement for near-zero maintenance costs.
In this paper we present Energy Efficient Light Routing in Utility Network
(EELRUN). This is a novel protocol that implements the routing without send-
ing extra control routing messages. This information (if needed) is combined
with ordinary resource information message. We have validate our protocol by
considering a network with white Gaussian noise and binary symmetric chan-
nels. The simulation shows a reduced fault while keeping a good throughput and
low power consumption.
2 Network
In our approach, each node has an error correction code table. The error correc-
tion code table of a node n assigns to each neighbour nf a value between zero
and one. This defines the error rate in the communication between nf and n.
The master node of the M-BUS has the error correction code tables of all nodes.
When a node modifies its error correction code table the new table is attached to
a resource information message in order to update the M-BUS controller tables.
In other words our protocol do not add extra control protocol messages to update
error correction code tables.
The M-BUS controller error code tables are initialised at the bootstrap phase
of the network or whenever a new slave is added. A slave announces its presence
using a hello packet. The reception of this packet triggers the initialisation of
the error correction code table. Figure 3a shows in detail the master behaviour
whenever a communication starts. The master sends a request packet in order
to receive the resource measured by a node. The packet structure is described in
Fig. 1 and contains the path to reach the destination node. The path is calculated
Energy Efficient Light Routing in Utility Network 747
by using Dijkstra algorithm where the edge connection between nodes is labelled
with the error rate. The master waits until a timeout is triggered or a response
is received.
Figure 3b shows the behaviour of a slave whenever a packet is received. A
slave can assume three kinds of roles: a listener, an intermediate or a destination
node. The former type listens to all packet in its neighbourhood and stores the
error correction code table using the Hamming error code (see Fig. 2). All the
nodes can play the listener role. The other roles instead are decided by the master
node. The intermediate node acts as a request forwarder in order to reach the
destination and whenever an intermediate node timeout occurs or a response
is forwarded back to the master together with the node correction table. The
destination node responds to the request packet using the received path and
provide back the measured resource value.
In a hop to hop communication, the path and the error correction code
table determines the dimension of the packet. When the packet size exceeds
the protocol limit, the usage of fragmentation is required. For simplicity in the
simulations, we consider a dynamic packet with no size restriction.
A packet is divided into a set of blocks. For each block, the neighbours
that retrieves the packet, computes the error correction code. Since we have no
physical layer assumption, the error correction code corresponds to the Hamming
error correction code and concerns three possible outcomes: received, damaged
but still recoverable and unrecoverable frame. The packet sender detects an
unrecoverable frame waiting the response from the next hop for a fixed amount
of time. The amount of time depends on the number of hops left to reach the
destination. This timeout feature is applied for request packets. Whenever a
timeout is triggered, the nodes of the routing path that can reach the master
provides back their error correction code table (if a variation occurs).
⎧
⎪
⎨0 ∀b ∈ B, h(b) = 0
ecc(B) = inf ∃b ∈ B, h(b) > 1 (1)
⎪
⎩ |{b∈B|h(b)=1}| otherwise
|B|
In the Eq. 1, we define the error correction code for a single packet as a function
of the set of blocks B using the Hamming function h. This returns the number
of bits in error. The cardinality of the codomain improves proportionally the
decision capability of the master node and as consequence the precision of the
master computed path.
Request <<abstract>> Response

Packet
+ command: String + success:Boolean
+ source : Node + errorCorrectionCodes : Integer[Node] {unique}
+ target : Node + destination : Node + payload: String
+ hopCount: Integer
+hamming() : Integer
+pathSize() : Integer
+removeHop(Integer)
Fig. 1. Packet types

Validate Packet
p:Packet
Unintelligible packet Readable packet
The error correction code
Update error correction

[p.hamming() < 2] Decode packet
code table
Fig. 2. Packet validation
<<singleton>>
Choose a recipient
req : Request
Compute the shortest
path
Create the request Validate Packet

packet with path
req : Request [Unintelligible packet]
[Readable packet] Unintelligible packet

hops = path
source = path[0]
destination = path[1] [req.destination <> self or req.target <> self]
hopCount = path.length
Listener node
[req.destination = self or req.target = self]
req.removeHop(self) req.removeHop(self)
Forward newReq : Request

req
target = req.target
hops = req.path
source = self
destination = req.path[0]
hopCount = req.hopCount - 1
Receive
response req.pathSize() * 2k
[req.target = self]
Target Create response

resp : Response communication with the node
failure error correction code table
and payload
Forward req.pathSize() * 2k
newReq
Validate newResp:Response
resp
success = true
source = self
Create response
destination = req.source Receive response with the node
Unintelligible noise table
[Unintelligible packet] packet
[Readable packet] Send childRes:Response createRes:Response

response
success = false
Update source = self
the graph weights destination = req.source
Validate Packet
Listener Reached
[resp.destination <> self] node [Unintelligible packet]
createRes
[Readable packet]
[resp.destination = self]
Add the node error
correction code table Hop response
Unintelligible packet in childRes timeout
[resp.success = false] [resp.success = true]
Send
childRes
Target not reached Communication Success
Hop response
(a) Master sending activity (b) Slave activity
Fig. 3. Network behaviour

2.1 Cache
In a master/slave communication, the message size affects the power consump-
tion of the node and also the number of unrecoverable messages. For each com-
munication, the master forwards a packet composed of header and payload. In
a smart meter system, the header part is bigger than the payload which is com-
posed of a routing path and a list of error correction code tables, whereas the
packet is a response type. In order to reduce the header part, we introduce a
cache system for optimise the routing path. In most of the distributed system,
the usage of caches has been deployed for different purposes [9]. This can also
be applied to routing. In our protocol, each slave has a cache table. The cache
table of a slave C assigns to a node cn the next hop. The master node has the
cache table of all nodes. As we can see from the Algorithm 2, the master do not
disseminates the routing path whereas all nodes know their next hop. On the
contrary the master includes the full path in the packet.
As we can see from the Algorithm 3, the master node introduces entries in the
cache when a communication returns. On the contrary, the Algorithm 1 shows the
removal of the cached path whether a fault occurs or a communication timeout.
Figure 4a shows the full master behaviour. Each slave updates their cache when
receives a packet with a path. Whenever the path is not provided, the node uses
the cache to get the next hop. Figure 4b shows the slave behaviour using cache.
3 Simulation
Seeing the complexity of the system, two simulations were executed for checking
the effectiveness of this protocol. In the former, the master provides the full path
whereas in the latter instead reduces the path through a distributed cache. Both
simulations require the following input parameters: messages - the number of
packets, nodes - the number of nodes in the network, connectivity - the number
of nodes reachable from each node, noise frequency - the number of packets
required for updating the noise, impact noise in the network- the percentage
of nodes affected by noise and noise edges impact - for each affected node, the
percentage of edges affected by noise. In simulations, the structure of the network
is randomly generated. The master previously chooses a list of unique slaves
and afterwards ask them cyclically the measurement. Due to having a reliable
simulation, a fundamental parameter is the number of messages. Although a high
number of packets are desirable, the simulation is not feasible for computation
reason.
In order to compute the Hamming correction code, we modelled the noise in
the network. For doing that, we introduced a weighted graph structure, described
by nodes and edges in which the weight of an edge is the noise level. During simu-
lations we alter the noise level in the network, considering a percentage of nodes.
For every node selected, the percentage of edges describes the number of edges
with a modified noise. These choices are made following a uniform distribution.
Moreover for computing the Hamming error, we consider the network as a binary
symmetric channel. Using such assumption, the Eq. 2 describe the probability of
<<singleton>>
Choose a
destination
req : Request
Compute the shortest

path Validate Packet
Reduce path
using cache [Unintelligible packet]
[Readable packet] Unintelligible packet
Create the request [req.destination <> self or req.target <> self]

packet with path
Listener node
[req.destination = self or req.target = self]
req : Request [req.hops->includes(self)

and self <> req.destination] and self <> req.destination]
hops = path [self = req.destination]

source = path[0]
destination = path[1]
next =
next = req.destination req.hops->indexOf(self)
related to req.target
req.removeHop(self)
<next, req.target>
inside the cache.
Forward
req
req.removeHop(self)
Receive
response req.getHopCount() * 2k
newReq : Request
target = req.target
hops = req.path
source = self
Clear cache destination = next
resp : Response for hopCount = req.hopCount - 1
destination
Target [req.target = self]

Validate communication
resp failure Create response
with the node
error correction code table
and payload
Unintelligible Forward
[Unintelligible packet] packet newReq
newResp:Response
[Readable packet] success = true Create response

source = self with the node
Receive response
destination = req.source error correction code
Update table
the graph weights
Send res:Response createRes:Response

response
[resp.destination <> self] success = false
Listener source = self
node Validate Packet destination = req.source
[resp.destination = self] Reached

[Unintelligible packet]
createRes
[Readable packet]
[resp.success = false] [resp.success = true] Append the error

correction code table Hop response
Unintelligible packet in res timeout
Clear cache Update cache

for with path Send
destination for res
destination
Hop response
Target not reached Communication Success
(a) Master sending activity (b) Slave activity
Fig. 4. Network behaviour using a cache system

Algorithm 1. Master Cache fault algorithm

procedure PathFault(path, caches)
last ← getDestination(path)
for cache ∈ caches do
for entry ∈ cache do
entry.recipients.remove(last)
end for
end for
end procedure
Algorithm 2. Master Path reduction algorithm

procedure PathReduction(path, caches)
list ← []
for (prev, next) ∈ path do
cache ← caches[prev]
if cache ∈ caches then Table cache hit
if (next, last) ∈ cache then Cache hit
list.insert(next, last)
end if
end if
end for
if list.length > 0 then Use the previous path.
return path
else Use an empty path
return []
end if
end procedure
Algorithm 3. Master Cache hit algorithm

procedure PathHit(path, caches)
for (prev, next) ∈ path do
cache ← caches[prev]
if cache ∈ caches then Table cache hit
cache.insert(next, last)
else Provide new cache
newCache ← [(next, (last))]
caches.insert(newCache)
end if
end for
end procedure
a recipient, using the noise level r and the message size n, to receive a readable,
a damaged but recoverable and an unrecoverable packet [8]. The noise level r is
a small value between 0 and 0,5. As a noise function, we used a white Gaussian
noise (with mu = 0.25 and sigma = 0.04) for describing a real communication
between nodes.
Using the noise level and the Eq. 2, we deduce the communication outcome
through the Hamming code applied to each block. The size of a block is 16 bit
where 11 bit are used for the payload and 5 bit are for Hamming redundancy.
(1 − r)n , (1 − r)n−1 ∗ r, 1 − ((1 − r)n−1 ·(r + (1 − r)n )) (2)

4 Results and Discussion

For computational reason, the simulations concern three kind of network: low,
medium and high noise. The Table 1 shows the network types parameters.
Table 1. Network types
Network Noise node impact Noise edge impact Noise frequency

Low 10% 10% 1%
Medium 25% 25% 5%
High 50% 50% 10%
The simulation output without cache optimisation is about the success rate
and the throughput average of the network. The success rate is the ratio between
the number of received messages and the number of messages sent by the master.
The definition of throughput average tha is the average of the throughput in a
single hop to hop communication. This parameter represents the number of bit
sent in one tick (the smallest time unit in our simulation) with a successful
communication.
100
Successrate
50
Low
Medium 10
8
6
Noise 4
High2 %Edges
100 100
Successrate
Successrate
50 50
Low Low
Medium 10 Medium 10
8 8
6 6
Noise 4 Noise 4
High2 %Edges High2 %Edges
Fig. 5. Success rate trend (100 nodes in orange, 500 nodes in blue and 1000 in red)
90
100
T hroughputindex
85
80
80
Low
10
8 75
Medium 6
4
Noise High2 %Edges
Fig. 6. Throughput index ti trend (red - 100 nodes, orange - 500 nodes, yellow - 1000
nodes)
For the cache simulation, we analysed only its throughput average thac. In
order to underline the cache impact, we define the throughput index ti as the
percentage of the ratio between the thac and tha.
Figure 5 shows the relation between parameters and the success rate. Such
correlations are not trivial and the simulations are a valid tool for evaluate them.
Figure 5 underlines the inverse proportional relationship between the success
rate and the noise impact. The protocol provides excellent success rate whether
the noise impact level is lower or equal to the medium one. Generally, the most
impact attribute is the network size together with the noise and the number of
edges. The communication path is influenced by all the factors. The number of
edges is not effective for lower noise level rather can be beneficial for the best
path. With greater level of noise, the edges affect more the network success rate,
especially considering bigger networks.
Figure 6 underlines the relationship between the network parameters and the
throughput index ti. As we can see, the network size holds an important role in
the throughput ti trend, which is also influenced by the number of edges. In most
of the case, the protocol does not provide any advantages using highly intercon-
nected networks. The reason concerns the minimum impact of the cache protocol
with short routing path. The cache system works well in a small network.
5 Conclusion
Our approach extends the Meter bus protocol with a noise resilient routing
protocol. This sends the path together with the message to be delivered. This
can cause a big overhead especially when the path is long. For this reason we
make use of a caching system. Our protocol performs well for medium and lower
noise level in a large networks. As future work, we plan to make a detailed energy
evaluation by using the model that is described [12].
References
1. Afzaal, A., Saleemi, M., Nazir, M., Saleemi, F.: Summarization of energy-efficient
communication protocols for smart grid. Int. J. Res. Rev. Comput. Sci. 169, 321–
354 (2018)
2. Saputro, N., Akkaya, K., Uludag, S.: A survey of routing protocols for smart grid
communications. Comput. Netw. 56, 11 (2012). https://doi.org/10.1016/j.comnet.
2012.03.027
3. Vannucchi, C., Diamanti, M., Mazzante, G., Cacciagrano, D.R., Corradini, F.,
Culmone, R., Gorogiannis, N., Mostarda, L., Raimondi, F.: vIRONy: a tool for
analysis and verification of ECA rules in intelligent environments. IEEE (2017)
4. Mohan, V.: An Introduction to Wireless M-Bus, Silicon Labs (2015)
5. DIN EN 13757-2: Communication systems for and remote reading of meters, Part
2: Physical and link layer. European Committee for Standardization (2005)
6. Communication systems for and remote reading of meters, Part 3: Dedicated appli-
cation layer. European Committee for Standardization (2005)
7. Data Commn and Networks (Isrd). Tata McGraw-Hill (2006). https://books.
google.it/books?id=6rRVfHq3hXoC
8. Bazlov, Y.: Coding Theory. Part 2: Hamming Distance (2010). www.maths.
manchester.ac.uk/∼pas/code/notes/part2.pdf
9. Paul, S., Fei, Z.: Distributed caching with centralized control. Comput. Commun.
(2010). https://doi.org/10.1016/S0140-3664(00)00322-4
10. Sikora, A., Werner, R., Grafmüller, J.O.: Design and implementation of an energy
aware routing extension for energy autarkic wireless M-Bus networks. In: Interna-
tional Conference on Advances in Computing, Communications and Informatics
(ICACCI) (2013)
11. Sikora, A., Werner, R., Grafmüller, J.O.: Design, implementation, and verification
of an energy autarkic, RF-based water meter with energy aware routing. In: Energy
self-sufficient Sensors; 7th GMM-Workshop (2014)
12. Ullah, Z., Mostarda, L., Gagliardi, R., Cacciagrano, D., Cacciagrano, D.: A Com-
parison of HEED based clustering algorithms - introducing ER-HEED. In: IEEE
AINA 2016, Crans-Montana, 23–25 March 2016
Innovative Technologies in Informatics
and Networking (WITIN-2019)
Performance of the 25 Gbps/100 Gbps
Fullmesh RoCE Network Using Mellanox
ConnetX-4 Lx Adapter and Ruijie S6500
Ethernet Switch
Hui Li(&), Xiaolong Chen, Tao Song, Haiyin Chen, and Hao Chen
Testing Group of the Data Center Network, Ruijie Network, Fuzhou, China
{huili,chenxiaolong,songtao,chy,chenh}@ruijie.com.cn
Abstract. Zero-copy RDMA is promising in the support of IPC operations for

a wide range of applications. Modern high-performance network protocols such
as Infiniband, RoCE, or iWARP offer RDMA, which provides high bandwidth
and low latency communication for applications. In this paper, we assess
25 Gbps/100 Gpbs fullmesh RoCE network using the Mellanox ConnetX-4 Lx
adapter and Ruijie s6500 switch, which are already largely deployed in Alibaba,
Baidu, and Tencent data center network environments. We find that the capa-
bility of the switch can be a bottleneck in the fullmesh network, and we also find
that our fullmesh deployment can achieve linear scalability of throughput as
long as the total throughput of servers does not exceed the link rate of switches.
1 Introduction
Computation, storage, and the web are the three main types of applications deployed in
data center or cloud. All these applications require IPC utilities to exchange interme-
diate data or control messages to process parallel tasks that cross over many servers. As
such, the overhead of moving data within the data center cannot be overlooked. Remote
direct memory access (RDMA) provides the ability of a device to access remote host
memory directly, without the intervention of the computer’s operating system.
The RDMA enables high-throughput, low-latency networking with low CPU utiliza-
tion, which is especially useful in parallel compute clusters.
The RDMA has the following major advantages: (1) Zero-copy, applications can
perform data transfer without the network software stack involvement and the data are
sent and received directly to the buffers without being copied between the network
layers. (2) Kernel bypass, applications that can perform data transfer directly from the
user space without the need to perform context switches. (3) No CPU involvement,
applications can access remote memory without consuming any CPU in the remote
machine. The caches in the remote CPU(s) will not be filled with the accessed memory
content.
There are several network protocols that support RDMA: (1) InfiniBand, a new
generation network protocol that supports RDMA natively from the beginning.
(2) RoCE, A network protocol that allows RDMA to be performed over Ethernet

https://doi.org/10.1007/978-3-030-15035-8_73
758 H. Li et al.
networks. (3) iWARP, A network protocol that allows performing RDMA over
TCP. In this paper, we are deploying Remote Direct Memory Access (RDMA) tech-
nology in fullmesh RoCE network to provide ultra-low latency and high throughput
network to applications, with very low CPU overhead.
Fig. 1. RDMA software stack
InfiniBand (abbreviated IB) is a computer-network communication standard used in

high-performance computing that give very high throughput and very low latency. It is
used for data interconnect both among and within computers. InfiniBand is also used as
either a direct or switched interconnection between servers and storage systems.
RoCE is the transport layer standard that provides for RDMA over an Ethernet
network. Unlike the InfiniBand, with RoCE you can use the same Ethernet cabling and
switching that you already have while it provides the capability of RDMA. There are
two RoCE versions, RoCE v1 and RoCE v2. RoCE v1 is an Ethernet link layer
protocol and hence allows communication between any two hosts in the same Ethernet
broadcast domain. RoCE v2 is an internet layer protocol which means that RoCE v2
packets can be routed. As depicts in Fig. 1, RoCEv2 retains the IB transport layer, but
replaces IB networking layer (L3) with IP and UDP encapsulation, and replaces IB L2
with Ethernet.
iWARP is an alternative RDMA offering that is more complex and unable to
achieve the same level of performance as RoCE-based solutions. iWARP uses a
complex mix of layers, including DDP (direct data placement), an adaptation known as
MPA (marker PDU aligned framing), and a separate RDMA protocol (RDMAP) to
deliver RDMA service over TCP/IP. This convoluted architecture is an ill-conceived
attempt to fit RDMA into the existing software transport framework. Unfortunately this
compromise causes iWARP to fail to deliver the three key benefits that RoCE is able to
achieve: high throughput, low latency, and low CPU utilization.
In network logical topology, communication devices are modeled as nodes and the
connections between the devices are modeled as lines between the nodes. Several types
of network logical topologies have been used in LANs, including ring, bus, mesh and
Performance of the 25 Gbps/100 Gbps Fullmesh RoCE Network 759
star. Figure 2 depicts a fullmesh with 11 nodes, which is a fully connected graph. In
Fig. 4, the nodes were connected to the switch in a bus topology. We also consider
Fig. 4 as a fullmesh network in logically. In Fig. 4, any two nodes are connected
through only 1 hopper or 2 hopper switch devices. As a result, the bandwidth and the
latency between any two nodes of a fullmesh network are optimized.
Fig. 2. Fullmesh with 11 nodes
This paper uses OFED, MPI, and Spark to evaluate the performance of
25 Gbps/100 Gbps fullmesh RoCE network. OFED refers to OpenFabrics Enterprise
Distribution. It is a package for Linux that includes all of the needed software (libraries,
utilities and drivers) to work with RDMA enabled protocols (Infiniband, RoCE, and
iWARP). It includes a PerfTest library that we will use to evaluate the performance of
25 Gpbs/100 Gpbs fullmesh RoCE network. The message passing interface (MPI) is a
standardized and portable message-passing standard designed by a group of researchers
from academia and industry. There are several MPI runtime implementations, which
include Intel MPI, MPICH, and OpenMPI. All three MPI implementations support the
RDMA protocol. We used the OpenMPI MPI_Alltoall API to evaluate the performance
of 25 Gbps/100 Gpbs fullmesh RoCE network. Spark is a unified analytics engine for
large-scale data processing. Spark powers a stack of libraries including SQL, Data-
Frames, and MLlib for machine learning. Mellanox sponsors the SparkRDMA project
which have Spark to use RDMA protocol with only several extra configuration work of
Spark. In this paper, we also use the SparkRDMA Terasort application to evaluate the
performance of 25 Gpbs/100 Gpbs fullmesh RoCE network.
The rest of the paper is organized as follows. We provide a brief overview of the
related work in Sect. 2. We illustrate the deployment architecture of the fullmesh
network in Sect. 3. In Sect. 4, we introduce four applications and evaluate the per-
formance of 25 Gbps/100 Gpbs fullmesh RoCE network. Finally, we conclude the
paper in Sect. 5.
760 H. Li et al.
2 Related Work
2.1 Next Generation Fullmesh Network
Paper [1] evaluated the performance of the RDMA over RoCE standard in an enterprise
data centers infrastructure. It also analyzed why the RoCE has received broad industry
support from many hardware, software and system vendors. The authors in paper [2]
discussed the latest development in RoCE congestion control. Hardware congestion
control reduces the latency of the congestion control loop; the loop reacts promptly to
congestion by throttling the transmission rate quickly and accurately. The authors also
surveyed architectural features that allow deployment of RoCE over lossy networks
and presented test results. Paper [3] designed a DSCP-based priority flow control
(PFC) mechanism to ensure large-scale deployment of RoCEv2. They addressed the
safety challenges brought by PFC-induced deadlock, RDMA transport livelock, and the
NIC PFC pause frame storm problem. Their experiments showed that the safety and
scalability issues of running RoCEv2 at scale can be addressed. Paper [4] describes the
challenges of using RDMA over commodity Ethernet (RoCEv2) to support some of
Microsoft’s highly-reliable, latency-sensitive services. Their experiences show that the
safety and scalability issues of running RoCEv2 at scale can all be addressed, and
RMDA can replace TCP for intra data center communications and achieve low latency,
low CPU overhead and high throughput.
2.2 Parallel Runtimes Using RDMA

Paper [5] focused on the redesign of the communication framework inside MapReduce.
They proposed the design of an RDMA-based Hadoop MapReduce over InfiniBand
and several design elements, such as data shuffle over InfiniBand. Their results show
that the RDMA-based solution achieves a performance improvement of 32% over
IP-over InfiniBand (IPoIB) and 21% over Hadoop-A. Paper [6] presented a novel
design of HDFS using RDMA over Infiniband via JNI interfaces. Experimental results
showed that, for 5 GB HDFS file writes, the new design reduces the communication
time by 87% and 30% over 1Gigabit Ethernet and IP-over-Infiniband, respectively, on
the QDR platform. Paper [7] focused on how RDMA and high-performance inter-
connects can benefit Spark. It utilized the plug-in based approach with the staged event-
driven architecture and RDMA-based designs provided both performance and pro-
ductivity. The performance results showed that Spark applications can benefit from an
RDMA enhanced design. Paper [8] explored the potential of employing RDMA to
improve the performance of online data processing (OLDP) workloads on MySQL
using Memcached for real-world web applications. This outcome revealed that
RDMA-enhanced Memcached can significantly improve the performance of OLDP
workloads that perform a sizeable number of concurrent reads compared to default
Memcached running with IPoIB QDR. Paper [17] rethink the widely used RPC
abstraction for network communication because the deep learning workloads
and network technologies such as RDMA are emerging. By designing a simple
“device”-like interface, along with a combination of static analysis and dynamic

tracing, they have enabled cross-stack optimizations for general deep neural network
training to take full advantage of the underlying RDMA capabilities.
2.3 Main Contribution of This Paper

One of the main contributions of this paper is to make a thorough performance eval-
uation of the 25 Gbps/100 Gbps fullmesh RoCE network using state-of-art hardware
devices. We also note several performance issues when using the fullmesh of RoCE
network for four state-of-the-art applications. The second contribution of this paper is
addressing how difficult it is to adapt RDMA to fullmesh applications. We show that it
does not require much programming effort to adapt RDMA to fullmesh applications
(Fig. 3).
Fig. 3. Intelligent stack
3 Deployment Architecture of 25 Gbps/100 Gbps Fullmesh

RoCE Network
Intelligent stack [9] (iStack) connects multiple switch devices into one switch device.
iStack is a nonstandard technology in DCN that is used to extend the ports of a switch.
Since it can simplify network management, iStack has been widely used in today’s
DCN network deployment. However, iStack requires that multiple switch devices must
come from the same vendor, and be installed with the same software, which conflicts
with the openness of the next generation DCN network. Nevertheless, it is easy to
dismantle the iStack if there is a fault between switch devices. The fault of one switch
easily affects the work of other switches.
VSU lite [10] technology can overcome the disadvantages of iStack while still
providing the advantages of iStack. VSU lite utilizes dual ports to uplink two switch
devices, and NIC is set as bond mode 4, which obeys the IEEE 802.3ad (RoCE LAG).
To make use of VSU lite, it requires modification of the network driver in the Linux
kernel in addition to some configurations in the switches. Figure 4 illustrates the
762 H. Li et al.
deployment architecture of 25 Gbps/100 Gpbs fullmesh RoCE network using VSU

lite. In Fig. 4, there are two groups of servers connected to 4 ASWs. The intragroup
message go through ASW, while the intergroup message go through ASW, PSW.
Priority flow control (PFC) is a recommended flow control mechanism that allows
for high RoCE performance and guarantees zero packet loss in the network. Initial
implementations of RoCE required a lossless network for optimal performance. This is
typically achieved by enabling PFC on Ethernet NICs and switches. Lossless networks
avoid dropping packets when the buffer is full by using PFC to pause the incoming
packet transmission before the buffers that receive the packets overfill. PFC comple-
ments Congestion Notification in Data Center Bridging networks. Operation of
priority-based flow control is limited to a domain controlled by a Data Center Bridging
control protocol that controls the application of Priority-based Flow Control, Enhanced
Transmission Selection, and Congestion Notification.
Fig. 4. VSU lite
For the best performance, we configure the PFC, ECN, DSCP, DCQCN services in
network adapters and switches for the RoCE network. Explicit congestion notification
(ECN) was initially defined for TCP/IP by embedding an indication for congestion in
the IP header and an acknowledgement in the TCP header. ECN compatible switches
and routers mark packets when congestion is detected. The congestion indication in IP
header is also used by congestion control of RoCEv2. Differentiated services code
point (DSCP) is an 8-bit field in the IP header. The network element will take the
priority from the DSCP bits and map it to the right buffer/queue/priority. DC-QCN is a
congestion control protocol for RoCE. It is based on the ECN feature on network
switches, to inform the sender to throttle the injection rate upon congestion.
Our testbed has16 Dell R720 servers. Each compute node has a 16-core 2.4 GHz
Intel Xeon E5-2630 processor with 16 GB of main memory, and it is equipped with
Mellanox ConnectX-4 LX (dual port 25 GbE/50 GbE) HCAs. We use CentOS Linux
7.4.1160 (Core) with kernel 3.10.0-693.2.2.el7.x86_64 with VSU lite patch. In addi-
tion, we use the Mellanox Open Fabrics Enterprise Distribution (OFED) [11]
MLNX_OFED_LINUX-4.2-1.0.0.
ConnectX-4 Lx EN is the adapter supports RoCE network produced by Mellanox.

It provides an unmatched combination of 1, 10, 25, 40, and 50 GbE bandwidth,
submicro second latency and 75 million packets per second message rate.
The RG-S6500 is an Ethernet switch produced by the Ruijie network. It is a high-
performance and high-density switch designed for a DCN network. It provides up to
48 25 Gbps ports and 8 100 Gbps ports. It has four 1 GHz CPUs, 2 GB RAM, and
4 4 MB cache.
4 Performance of 25 Gpbs/100 Gpbs Fullmesh RoCE

Network
4.1 Performance of Point to Point Using MPI PingPong

We make two sets of experiments to evaluate performance of RDMA of point to point
operations by using Intel’s MPI PingPong jobs [12]. The first set of experiment is the
s-s pattern, which means that the server connects to the server directly with no switch.
The second set of experiment is the s-sw-s pattern, which means that two servers are
connected through a switch.
In the two set of experiments, the PingPong test sends and receives small message
for 1000 times using MPI sends/recv primitives. Figure 5 shows that the 25 Gbps
s-s pattern and 10 Gpbs s-s pattern can provide stable throughput performance for
message sizes varying from 1 byte to 134 MB, while the 25 Gpbs s-sw-s pattern and
10 Gpbs s-sw-s pattern have very serious performance degradation when message sizes
are larger than 5 KB. There are several reasons for the performance degradation. The
first reason is network congestion caused by the buffer overflow of the RG-S6500-4c
switch. In other word, the RG-S6500-4c switch does not have sufficient buffer to
transfer the data sent out by Mellanox ConnectX-4 Lx adapter. Second, we did not
configure the priority flow control (PFC) for this experiments, which results in a
significant performance penalty for network congestion.
Fig. 5. Bandwidth of Intel MPI PingPong job

764 H. Li et al.
4.2 Performance of Fullmesh Using OFED ib_write_bw

We made the performance experiments of 25 Gpbs/100 Gpbs fullmesh RoCE network
using OFED scripts with up to 16 physical servers. In this experiment and all the
experiments following this one, the topology of experiments is shown in Fig. 4. There
were two groups of servers, each of which had 8 physical servers. In each group, the
physical servers were connected to 2 dedicated ASW switches with VSU lite enabled.
Messages between two groups had to go through PSW as shown in Fig. 4. We use 8
servers to mimic single-rack traffic, and use 16 servers to mimic inter-rack traffic.
Figure 6 illustrates the pseudo code of the fullmesh of OFED scripts for both the client
and server functions. It has only 30 code lines. In the pseudo code, we spawn N-1
server processes and N-1 client processes on each server to mimic the fullmesh
communications of N * (N-1) connections. We also set the number of queue pairs to
200 to increase bandwidth utilization for data transferring. In addition, there is only
little performance fluctuations in bandwidth when setting up the queue pairs as 200.
Figure 7 shows the accumulated throughput of send and receive of OFED
ib_write_bw scripts using up to 16 physical servers. The bandwidth of single server is
more than 45 Gbps, and this is because we use RoCE LAG technology which leverages
the full capability of ConnectX-4 Lx adapter. The throughput is shown to be stable and
linearly scalable, the throughput is linearly scalable with the number of physical ser-
vers. The results indicate that the total throughput over 16 physical servers, which is
650 Gpbs, did not exceed the link rate of ASW and PSW. As a result we do not see the
performance bottleneck of throughput of fullmesh network with the 16 nodes. We
should note that it may show the performance bottleneck of fullmesh network when the
total throughput of all nodes is more than 800 Gbps, which is the link rate of PSW in
Fig. 4.
Fig. 6. Throughput of fullmesh network using OFED scripts with up to 16 nodes

4.3 Performance of Fullmesh Using MPI Alltoall

MPI_Alltoall is an MPI collective operation [13] in which all processes send the same
amount of data to each other, and receive the same amount of data from each other.
Each process breaks up its local sendbuf into n blocks. Process j sends the k-th block of
its local sendbuf to process k. The amount of data sent must be equal to the amount of
data received.
Figure 7 is the pseudo code of invoking MPI_Alltoall. The code is very simple, and
it has only 21 code lines. We set sendcount as 1024 * 1024 to have high network
throughput. We set the repeat number to 1000 to make the MPI_Alltoall primitives run
1000 times so as to have sustainable high network throughput.
Figure 7 shows the total throughput of fullmesh network using MPI_alltoall over
up to 16 nodes. The throughput is shown to be stable. It also shows linear scalability of
throughput using up to 16 nodes. The average throughput of each server node is nearly
25 Gbps, which is approximately half the 50 Gbps, the link rate of ConnectX-4Lx
adapter. The reason is that we use the rdma_cm lib of OFED for this MPI program,
which does not use the capability of the RoCE LAG. As a result, it only leverages half
capability of ConnectX-4 Lx adapter.
4.4 Performance of Fullmesh Using Spark Terasort

Terasort [14, 16] is an important benchmark that measures the amount of time to sort
one terabyte of randomly distributed data on a given computer system. SparkRDMA
Terasort is the Terabyte sort program run on SparkRDMA [15]. We have 3 Spark
applications to run the SparkRDMA Terasort applications:
1. TeraGen is a map/reduce program to generate the data.
2. TeraSort samples the input data and uses map/reduce to sort the data into a total
order.
3. TeraValidate is a map/reduce program that validates the output is sorted.
Terasort is a standard map/reduce tasks, except a custom partitioner. The custom
partitioner uses a sorted list of N-1 sampled keys that define the key range for each
reduce. In particular, all keys such that sample[i-1]<=key<sample[i] are sent to reduce
i. It guarantees that the output of reduce i are all less than the output of reduce i+1. To
speed up the partitioning, the partitioner builds a two level trie that quickly indexes into
the list of sample keys based on the first two bytes of the key. TeraSort generates the
sample keys by sampling the input before the job is submitted and writing the list of
keys into HDFS.
766 H. Li et al.
Fig. 7. Throughput of fullmesh network using MPI_Alltoall over up to 16 nodes.
We ran the Terasort code on 16 nodes and sorted the 100 million records (10 GB)
within 5 min. Figure 8 show the total throughput of fullmesh network using Spark
Terasort with up to 16 nodes. The throughput is shown to be stable. It also shows the
linear scalability of the total throughput using up to 16 server nodes. However, we see
slightly fluctuations of throughput when number of nodes are 2, 3, 9, and 13. This is
because the input data is a little skewed so we could see the unfair throughput. The
result of the average throughput of each node is nearly 300 Mbps; this value is very
small because the disk I/O operations are involved in the shuffle stage of the Spark
Terasort task. In fact, even considering the peak performance of throughput of Terasort
Spark task, which is about 2 Gbps, and the total peak throughput is 2 * 16 = 32 Gpbs,
which is still much lower than the link rate of the switch.
Fig. 8 Throughput of fullmesh network using Spark Terasort over up to 16 nodes.

5 Summary and Conclusion
This paper depicts the deployment architecture of 25 Gpbs/100 Gpbs fullmesh RoCE
network using VSU-lite. This paper studies the performance of throughput of
25 Gpbs/100 Gpbs fullmesh RoCE network using four applications, the MPI PingPong
jobs, the OFED scripts, MPI Alltoall primitives, and SparkRDMA Terasort. The results
show that the capability of switch and performance of disk I/O could be a performance
bottleneck in the fullmesh network. This paper also shows linear scalability of total
throughput of fullmesh network using up to 16 nodes as long as the total throughput
does not exceed the link rate of switch. We believe that our deployment architecture
and configurations of 25 Gpbs/100 Gpbs fullmesh RoCE network are optimal and can
be widely used in academia and industrial environment.
References
1. Beck, M., Kagan, M.: Performance evaluation of the RDMA over Ethernet (RoCE) standard
in enterprise data centers infrastructure. In: Proceedings of DC-CaVES (2011)
2. Guo, C., Wu, H., Deng, Z., Soni, G., Ye, J., Padhye, J., Lipshteyn, M.: RDMA over
commodity ethernet at scale. In: SIGCOMM 2016 (2016)
3. Shpiner, A., Zahavi, E., Dahley, O., Barnea, A., Damsker, R., Yekelis, G., Zus, M., Kuta, E.,
Baram, D.: RoCE Rocks without PFC: Detailed Evaluation
4. Zhu, Y., et al.: Congestion control for large-scale RDMA deployments. In:
ACM SIGCOMM (2015)
5. Islam, N.S., Rahman, M.W., Jose, J., Rajachandrasekar, R., Wang, H., Subgramoni, H.,
Murthy, C., Panda, D.K.: High-performance RDMA-based design of Hadoop mapreduce
over Infiniband. In: Proceedings of the International Conference on High Performance
Computing, Networking, Storage and Analysis, SC 2012 (2012)
6. Lu, X., Islam, N.S., Wasi-Ur-Rahman, M., Jose, J., Subramoni, H.: High-performance
design of Hadoop RPC with RDMA over InfiniBand. In: International Conference on
Parallel Processing (2013)
7. Accelerating Spark with RDMA for Big Data Processing: Early Experiences
8. Shankar, D., Lu, X., Jose, J., Wasi-ur-Rahman, M., Islam, N., Panda, D.K.: Can RDMA
benefit online data processing workloads on memcached and MySQL. ISPASS, Philadel-
phia, PA, USA (2015)
9. Stack Configuration – Huawei Technical Support. http://support.huawei.com/enterprise/
docinforeader!loadDocument1.action?contentId=DOC1000074401&partNo=10042
10. How to implement the VSU-lite in DCN network. http://www.ruijie.com.cn/fa/xw-hlw/
61546
11. https://www.mellanox.com/related-docs/prod_software/Mellanox_OFED_Linux_User_
Manual_v4.0.pdf
12. Mellanox OFED for Linux User Manual
13. https://software.intel.com/en-us/articles/intel-mpi-benchmarks
14. https://www.open-mpi.org/doc/v1.10/man3/MPI_Alltoall.3.php
15. https://github.com/ehiggs/spark-terasort
16. https://github.com/Mellanox/SparkRDMA
The Warning System for Speed Cameras
on the Road by Deep Learning
Chien-Chung Wu1(&), Yu-Xuan Lin1, Deng-Xiang Hu1,

Chien-Chuan Ko2, and Ji-Han Jiang3
1
Southern Taiwan University of Science and Technology, Tainan, Taiwan
wucc@stust.edu.tw
2
National Chiayi University, Chiayi, Taiwan
kocc@mail.ncyu.edu.tw
3
National Formosa University, Yunlin, Taiwan
jhjiang@nfu.edu.tw
Abstract. In order to reduce traffic accidents, government authorities would

install speed cameras for identifying vehicles travelling over the legal speed
limit. However, drivers often ignored warnings and were handed tickets, and
meanwhile it did not achieve the purpose of setting up the cameras. This paper
took the road in Taiwan as an example. By installing a camera in front of the
vehicle, it could capture the image of the road ahead. This system was imple-
mented on the NVIDIA Jetson TX2 and used YOLO V3 as the architecture for
CNN classification. The images were divided into six categories: “police patrol
car”, “warning sign for front speed camera”, “warning slogan for front speed
camera”, “front view of radar speed camera”, “back view of radar speed cam-
era”, “traffic sign of speed limit”. When the first five categories were detected,
the system would issue a warning, and when “the traffic sign of speed limit” was
detected, the number in the middle part would be identified to obtain and update
the speed limit. The result of the experiment on the YOLO V3 when outputting
6 categories was that Mean AP was 81.25%, and Average Recall was 85.5%.
The result of the digital recognition for speed limit had a precision of 83.5% and
a recall of 80.2%.
Keywords: Convolutional neural network Embedded system

Speed limit detection Speed cameras alert YOLO CNN
1 Introduction
According to the statistics of 2002–2016 from the Taipei City Government Police
Department, the traffic accidents still accounted for the sixth place in the top ten causes
of death, and the percentage of traffic accidents was the highest. The causes of traffic
accidents include speeding, drunk-driving, zigzag-driving and driving with negligence.
However, in fact, when driving, due to advances in vehicle manufacturing tech-
nology, the driver may unconsciously speed up. It may also be because the driver is not
familiar with the road or terrain, leading to speeding unconsciously when going
downhill. Due to these violations, drivers may get tickets and are more likely to have
car accidents. Though the speed limit guidelines for roads are clearly stated in traffic
https://doi.org/10.1007/978-3-030-15035-8_74
The Warning System for Speed Cameras on the Road by Deep Learning 769
regulations, in addition to driving on highways or roads with specific speed limits, the
speed limits on ordinary roads often differ in different sections. The information of
different road sections often make drivers confused with the legal speed limit It is very
dangerous as drivers hit the brake when they are watching a speed camera or a police
car. This paper provides image capture by setting up a camera in front the car, and
provide the information of the speed limit sign to the driver through image classifi-
cation and recognition.
When the “police patrol car”, “warning sign for front speed camera”, “warning
slogan for front speed camera”, “front view of radar speed camera”, “back view of
radar speed camera”, “traffic sign of speed limit” were detected, the system would issue
a warning, and when “the traffic sign of speed limit” was detected, the number in the
middle part would be identified to obtain and update the speed limit.
Wu et al. [1] integrated the Haar-like algorithm to filter out possible detection
targets, and then used the CNN classifier to calculate the possible classification cate-
gory scores of the target, and then determined whether it was the probability of the
target and indicated the position. Yavas et al. [2] installed RFID tags in a wide range on
the traffic infrastructure, and installed RFID readers on the vehicles. When the vehicles
passed the RFID tag, they read pre-existing speed limit data in the tag.
Moreover, it has been a hot research topic in recent years on image recognition and
classification through Convolutional Neural Networks (CNN). Luo et al. [3] used the
multi-task convolutional neural network to identify traffic signs. Maurice Peemen et al.
[4] used the operation of convolutional neural networks to extract features to classify
visual items. For applications in portable and wearable devices and using hardware
accelerators to achieve and overcome the burden of computing, it could be used to
identify road speed limits on the road.
2 Design of System
As mentioned above, the paper used the embedded system to realize the warning
system for speed cameras on the road through image capture and deep learning tech-
nology. The system analyzed and identified speed limit traffic signs in real time through
image recognition technology and recorded them. While driving, “the police patrol car
on the roadside”, “the radar speed camera” and “the warning sign of front speed
camera” were recognized, the system would automatically release the warning message
through the voice and inform the maximum speed of the road.
In order to achieve the function, the hardware of the system mainly adopted
NVIDIA Jetson TX2, the CPU adopted dual-core Denver 2 64-bit CPU quad-core
ARM® A57 Complex, GPU adopted NVIDIA PascalTM architecture and 8 GB L128
bit DDR4 memory. Besides, the messages from the Bluetooth OBD-II decoder were
received via the USB Bluetooth dongle, mainly the speed of the vehicle. The images in
front of the vehicle were connected to the system via USB via SJCAM’s SJ4000 sport
camera. The diagram is shown in Fig. 1.
The development of the system was divided into training host and the embedded
vehicle implementation platform. The main training work required a lot of computing
power, therefore it had been trained on the host computer, the weight file of the trained
770 C.-C. Wu et al.
Fig. 1. System architecture diagram
neural network was transplanted to the embedded vehicle implementation platform.

The next is to explain the entire development process.
Because Taiwan’s traffic signs, radar speed cameras and traffic service police
vehicles are different from foreign countries [5], the training patterns of this system
were collected from recording videos on dashboard cameras, and the films shared by
other people on YouTube. To analyze each frame by the sample-collecting tool
LabelImg [6], and to perform sample target image capture by manual capture for
manual labeling and classification, the captured images and the recorded files were
respectively stored in the relevant category directory. After the training file had been
set, the training started. After the error reached the expected level and the set training
times were executed, the weighted file was to be stored, and then the test sample was
used for testing. If the result failed to meet the expectation, it would be revised and
trained again. The flow chart is shown in Fig. 2. The main training was on the Ubuntu
16.04 mainframe installed with NVIDIA GeForce 1070.
Fig. 2. Schematic diagram of the training process
As for the Convolutional Neural Networks, in order to speed up the object

recognition, this paper used YOLO [7] as the neural network architecture of real-time
object detection, mainly using Taiwan’s roads as an example, and classifying by
camera image input. The samples were collected from 30 different videos which were
self-captured or downloaded from YouTube. The classification of the output was

divided into six categories, including “police patrol car”, “warning sign for front speed
camera”, “front view of radar speed camera”, “back view of radar speed camera”,
“Traffic sign of speed limit”. The classification diagram is shown in Fig. 3.
Fig. 3. The classification diagram
Next, when the training had been completed, the original file of darknet would be
directly modified. When judging that the output category was “traffic sign of speed
limit”, the image of “traffic sign of speed limit” would be transmitted to the digital
recognition program which was added to the system for digital identification. The
program only took hundreds and units digits for digital recognition, and finally mul-
tiplied the identification result by 10.
To recognize the numerical value of the highest speed limit correctly is very
important in this system. However, the same speed limit has different meanings in
different lanes. For example, when seeing a speed limit sign of 50 on the highway, if
the vehicle is driving on the main road of the highway, this speed limit sign cannot be
misidentified by the system unless it is to leave the interchange, as shown in Fig. 4(a)
(white circle frame). Conversely, if the vehicle is leaving the interchange to exit the
highway, the 50 speed limit sign would be meaningful, as shown in Fig. 4(b). In order
to avoid this problem, the system deliberately avoided relatively small samples when
collecting training samples. At the same time, in order to avoid recognition errors when
identifying the sign of speed limit as well as avoid misjudging the speed limit of
different lanes, the detected category would have to be in 30 30 size for digital
recognition. The system would ignore it when the detected traffic sign of speed limit
was less than the size.
Next, the weights files of Convolutional Neural Networks trained by the host had to
be ported to the embedded system. This system used NVIDIA Jetson TX2 to imple-
ment the host, the system was installed Ubuntu 16.04, and darknet had to download
source code [8]. The Digital recognition program had to be recompiled through the
development tools [9], and the trained weights file had to be copied to the NVIDIA
Jetson TX2 system. The installation architecture is shown in Fig. 5.
772 C.-C. Wu et al.
Fig. 4. The speed limit sign of leaving the interchange, (a) the screen on the main lane, and
(b) the screen when leaving the interchange
Fig. 5. System porting diagram
3 The Results
The parameter for training convolutional neural networks was to set 64 images of
“batch” to find features. The maximum number of iterations was set to 40,000 times.
The initial learning rate of convolutional neural network was 0.001, and after the
20,000th Batch, the learning rate was reduced 0.1 times. The final learning rate was
0.00001.
Besides, the system was mounted in the lab, as shown in Fig. 6(a). The SJ4000
sport camera was connected via USB interface to provide real-time image input of the
system. The lens position could be controlled in the laboratory through a two-axis
steering platform and connected to the Bluetooth OBD-II via USB Bluetooth Dongle
and Buzzer via GPIO. The photo of the system installed in the car is shown in Fig. 6
(b). The system used a cigarette lighter interface to provide a voltage regulator circuit
with a voltage of 12 V, which then provided battery charging and provided power to
the entire system. In the lower right corner is the photo editing of the Bluetooth OBD-II
installed under the driver’s seat. There was not enough space in the front seat, so the
two-axis steering platform was removed.
This paper is mainly to provide the speed limit for the road section while driving. It
can also provide warnings when it detects police patrol cars, radar speed cameras and
warning signs. In addition to avoiding receiving the ticket, it can avoid overspeeding on
dangerous roads. However, there was no LCD display in NVIDIA Jetson TX2, the
Fig. 6 (a) Photo of the system mounted in the laboratory, (b) photo of the system installed
in the car
following result was captured in host computer. The detection result in the urban area is
shown in Fig. 7(a). The speed limit of 60 km/hr. and the speed camera were detected in
the screen, and the probability of displaying the identification result were 82% and 71%
respectively. Figure 7(b) shows the result of detection of “warning slogan for front
speed camera” and “warning sign for front speed camera” in the urban area, and the
probability of displaying the identification result were 76% and 68% respectively.
Fig. 7. Image of identification result, (a) the detection of the sign of speed limit and speed
camera in the urban area, (b) the result of detection of warning sign and warning slogan
In deep learning, the completed training model required an objective calculation

method to assess the quality of the model to ensure that the result of the training
contributed to the improvement of the model identification ability. Average Precision
(mAP) is the main indicator for evaluating the training model in this dissertation. As
represented by formula (1), it is mainly to calculate the product of precision and recall.
N is the total number of positive sample images.
X
N
mAP ¼ PðkÞ DrðkÞ ð1Þ
k¼1
774 C.-C. Wu et al.
The relationship between Avg_Recall and count is shown in formula (2). Recall
was defined as the ratio which was identified to be correct in the positive sample, while
count was the total number of inputs of positive samples.
DetectorNum
Avg Recall ¼ ð2Þ
count
There were 15 test videos collected, including 9 videos in urban areas and 6 videos
on the highway, which were self-captured. The result of the experiment on the YOLO
V3 when outputting 6 categories was that Mean AP was 81.25%, and Average Recall
was 85.5%. The result of the digital recognition for speed limit had a precision of
83.5% and a recall of 80.2%.
4 Conclusion
In this paper, the experiment was completed with NVIDIA Jetson TX2 and used
YOLO V3 as the architecture for CNN classification. The images were divided into six
categories: “police patrol car”, “warning sign for front speed camera”, “warning slogan
for front speed camera”, “front view of radar speed camera”, “back view of radar speed
camera”, “traffic sign of speed limit”. The result of the experiment on the YOLO V3
when outputting 6 categories was that Mean AP was 81.25%, and Average Recall was
85.5%. The result of the digital recognition for speed limit had a precision of 83.5%
and a recall of 80.2%.
Acknowledgments. This work was supported by the Taiwan Ministry of Science and Tech-
nology MOST 107-2221-E-218-023-MY2.
References
1. Wu, C.C., Wu, S.L.: A novel method for traffic sign detection and recognition. In:
Proceedings of the 2017 International Symposium on Novel and Sustainable Technology,
pp. B-12–B-13 (2017)
2. Yavas, B., Ozbilen, M.M., Yilmaz, Y.: Traffic speed sign recognition with RFID support. In:
Proceedings of the 2014 IEEE 8th International Conference on Application of Information
and Communication Technologies, pp. 1–4 (2014)
3. Luo, H., Yang, Y., Tong, B., Wu, F., Fan, B.: Traffic sign recognition using a multi-task
convolutional neural network. IEEE Trans. Intell. Transp. Syst. 19, 1–12 (2017)
4. Peemen, M., Shi, R., Lal, S, Juurlink, B., Mesman, B., Corporaal, H.: The neuro vector
engine: flexibility to improve convolutional net efficiency for wearable vision. In: Proceedings
of the 2016 Design, Automation & Test in Europe Conference & Exhibition, pp. 1604–1609
(2016)
5. Road Traffic Safety Portal Site in Taiwan. http://168.motc.gov.tw/News_Photo.aspx?n=
n1oOkV$gy2YXqPriGaYUcQ@@&sms=Qck7tL6R3WJmM6srNWCCKA@@
6. Tzutalin, D.: labelImg GitHub site. https://github.com/tzutalin/labelImg
7. Redmon, J., Divvala, S., Girshick, R., Farhadi, A.: You only look once: unified, real-time
object detection. In: Proceedings of the IEEE Conference on Computer Vision and Pattern
Recognition, pp. 779–788 (2016)
8. Redmon, J.: Darknet GitHub site. https://github.com/pjreddie/darknet
9. Redmon, J.: Installing Darknet website. https://pjreddie.com/darknet/install/
Using Feature Selection to Improve
Performance of Three-Tier Intrusion Detection
System
Yi-Jen Su1(&), Pei-Yu Huang1, Wu-Chih Hu2, Hsuan-Yu Lin3,

Chen-Yu Kao3, Shan-Hsiung Hsieh3, and Chun-Li Lin4
1
Shu-Te University, Kaohsiung City, Taiwan
{iansu,s17639115}@stu.edu.tw
2
National Penghu University of Science and Technology, Penghu City, Taiwan
wchu@npu.edu.tw
3
Telecom Technology Center, Kaohsiung City, Taiwan
{hylin,calos.kao,hsiung}@ttc.org.tw
4
National Center for Cyber Security Technology, Taipei City, Taiwan
cllin@nccst.nat.gov.tw
Abstract. Social media services have become an essential part of daily life. Once
5G services launch in the near future, the annual network IP flow can be expected
to increase significantly. In case of security threats, network attacks will become
more various and harder to detect. The intrusion detection system (IDS) in the
network defense system is in charge of detecting malicious activities online. The
research proposed an intelligent three-tier IDS that can process high-speed net-
work flow and classify attack behaviors into nine kinds of attacks by seven
machine learning methods. Based on the operation time, the detection process can
be divided into the offline phase, which trains models by machine learning, and
the online phase, which enhances the detection rate of network attacks by a three-
tier filtering process. In the experiment, UNSW-NB15 was adopted as the dataset,
where the accuracy of intrusion detection approached 98%.
1 Introduction
With the recent surge of network attack events, intrusion detection has already become
one of the most important technologies to enhance the security of computer systems.
The intrusion detection system (IDS) was first proposed by Anderson [1] in 1980.
Denning [2] also applied expert system in IDS applications in 1987 and monitored user
behaviors to identify anomalous operations. As information technology advances,
researchers used machine learning with feature selection to develop IDS for improved
pattern recognition. Hall [3] adopted correlation-based feature selection with machine
learning to remove irrelevant, redundant and noise features. In 2002, Mukkamala et al.
[4] applied the neural network and Support Vector Machines (SVM) on IDS design.
Next, they merged important features selection, neural network, and SVM on IDS
construction in 2003 [5]. Kumar et al. [6] compared three feature selection methods

https://doi.org/10.1007/978-3-030-15035-8_75
Using Feature Selection to Improve Performance 777
(correlation-based, information gain, and gain ratio) to effectively promote the accuracy
of the naïve Bayes classifier. They also proved that the feature selection method can
really optimize the machine learning classifier [7–9].
The most important feature of IDS is that it detects malicious network behaviors
from real network flow and automates related processes to protect the whole computer
system in time. The major IDS detection methods can be divided into three categories,
including misuse detection [10], anomaly detection [11], and hybrid detection [12, 13].
Misuse detection, a signature-based method, has higher accuracy than anomaly
detection, but cannot detect unknown attack behaviors. Both statistical analysis and
data mining are adopted to create a classification model for anomaly detection. The
advantage of anomaly detection is that it can detect unknown attacks, but its detection
accuracy is lower than misuse detection. Hybrid detection not only has the advantage of
the previous two detection methods but can also detect unknown network attacks.
Based on Cisco Visual Networking Index (VNI) [14] prediction, the global IP flow
will reach 4.8 ZB per year in 2022. In 2018, 3GPP also launched the first specification
of 5G and the data transfer rate of 5G is 20 Gb per second. Merging 5G and IoT, the
industry will change sharply. 5G technology will for sure bring people more conve-
niences in life, but information security works will also face extreme challenges. To
maintain high efficiency and accuracy in a high-speed network environment, this
research proposed an intelligent IDS to detect real-time network attacks. All operations
of the IDS can be divided into an offline phase and an online phase. The offline phase
adopts (1) machine learning methods to train known attack models and (2) the feature
selection method to optimize identification results. The online phase identifies mali-
cious behaviors in real-time in three-tier works. The first tier uses both a black list and a
white list to directly filter out either “normal” or “abnormal” connections with the
known websites. The tier operation can remove large amounts of network flow to skip
the detection that follows. In the second tier, the trained attack patterns are adopted to
determine whether the network flow is normal or abnormal. Only abnormal behaviors
will be sent to the third tier to check which kind of attack the abnormal behaviors
belong to. Finally, all intrusion detection results will be showed on a GUI to indicate
the current health status of the network environment.
2 Related Works
2.1 Machine Learning

Traditionally, people manually insert rules to the knowledge base to implement expert
systems. But machine learning, a subset of Artificial Intelligence (AI), aims to construct
an intelligent system by mining patterns/rules from large amounts of data. Most of the
time, some algorithms are adopted to enable the system to “self-learn” for better
778 Y.-J. Su et al.
prediction results. Basically, machine learning can be divided into the following
4 categories:
• Supervised Learning: The dataset is labeled to learn data patterns. Once new data is
input, the system can use these data patterns to predict the label of the input data.
This kind of learning algorithms includes SVM, Naïve Bayes, Decision Trees,
Linear Regression, Logistic Regression and so on.
• Unsupervised Learning: Without a labeling process for data training, the machine
learning method cannot judge the accuracy of classification result. In the training
process, as soon as cases are inserted into the machine, the system can automatically
find potential rules from the input dataset. This kind of machine learning methods
includes K-Means, Hierarchical classification and so on.
• Semi-Supervised Learning: The learning method is between supervised learning
and unsupervised learning. The method uses some labeled data and large amounts
of unlabeled data for the learning work. It saves the cost of manually labeling
everything and improves the accuracy of unsupervised learning.
• Reinforcement Learning: The agents select proper actions based on observation.
They evaluate feedback rewards to accumulate experiences to reach the biggest
reward from long-term actions.
2.2 Feature Selection

Recently, feature selection has become a key technology in enhancing IDS perfor-
mance. Using feature selection to filter out critical features of network flow behaviors
can reduce comparison time in the offline process of model training and in the online
process of malicious behavior detection. Comparing all features not only increases the
computation cost but also lowers the error rate in classification. The purpose of feature
selection is to maintain high detection accuracy and speed up the identification of
machine learning. Basically, feature selection methods can be divided into the followed
three kinds: Filter, Wrapper, and Embedded [15].
• Filter: Some measurement methods, including Relief, Correlation-based Feature
Selection and so on, are adopted to evaluate the relevance value between features
and labels and remove low relevance features. The computation cost of the filter
method is lower than the cost of the wrapper method; therefore, filter methods are
more suitable for datasets with a large number of features.
• Wrapper: The wrapper method trains models for training sets. It uses an induction
algorithm to search feature subsets and then applies learning algorithms to evaluate
features or feature subsets, In general, the wrapper method has higher computation
cost, but it can generate the best feature subset.
• Embedded: As the wrapper method uses learning algorithms to complete feature
selection, the embedded method incrementally selects features in the training pro-
cess to reduce the training cost of each feature set. Therefore, the computation cost
of the embedded method is lower than the wrapper method.
2.3 UNSW-NB15
In 2015, the Australian Defense Force Academy (ADFA) of University of New South
Wales (UNSW) Australia published an intrusion detection dataset named UNSW-
NB15 (University of New South Wale Network Based 2015) [16]. The total amount of
data in UNSW-NB15 is 2,540,044 records. Each data record contains 47 number of
features and 2 labels. One label shows that the network flow behavior is either normal
or anomalous, while the other one stands for the attack types. Basically, these 47
features are divided into 5 subgroups including Flow features, Basic features, Content
features, Time features, and Additional generated features. In addition, there are 9 kinds
of attack types in the UNSW-NB15 dataset including Analysis, Backdoors, DoS,
Exploits, Fuzzers, Generic, Reconnaissance, Shellcode, and Worms [16].
3 Research Method
The research proposed a three-tier intrusion detection system for the recognition of
malicious behaviors from real network flow. To reach higher accuracy, machine
learning and feature selection were adopted to detect attack behaviors, as shown in
Fig. 1. Based on the operation time, the system could be divided into two operations,
the offline phase and the online phase.
Fig. 1. System architecture of three-tier IDS
The online phase adopted a three-tier filtering mechanism to speed up intrusion

detection and reduce part of the records in each tier. In the first tier, an in-line network
probe was adopted to complete the first-tier filtering with a black list and a white list.
Some well-known IP addresses or domain names were labeled “safe” or “unsafe”.
780 Y.-J. Su et al.
IDS directly removed these network flows from the checking list, e.g. Facebook,
Google, YouTube and so on. The first-tier filtering process can remove lots of highly
dangerous/safe network flow and largely reduce the detection load of the second tier.
Before the detection work, the second-tier loaded 7 kinds of machine learning models
trained in the offline phase. If all trained models cannot find any attack behavior from
the current connection, the connection directly passed the third-tier checking. Once the
training models voted that the current connection was an attack, the connection would
be sent to the third-tier to identify the attack.
In the offline phase, the proposed IDS adopted machine learning methods to train
attack patterns for optimizing online phase operations. These trained models were used
to detect the network attacks. The related operations are shown in Fig. 2.
Fig. 2. Operation of the offline phase of IDS
The main purpose of feature selection is to speed up the pattern recognition of

network attacks by the IDS. Therefore, feature selection is very critical in optimizing
intrusion detection. It not only speeds up the detection by pattern matching, but also
maintains the detection accuracy. Weka 3 [19], a famous open source data mining tool
developed by Waikato university, provides data processing functions including
preparation, classification, regression, clustering, association rules mining, feature
selection, and visualization.
There are 8 kinds of feature selection methods distributed in the filter and wrapper
categories [20] adopted in this research. Weka 3 provides the feature selection methods
including CFS (Correlation-based Feature Selection), Correlation, Information Gain,
Gain Ratio, OneR, RelifeF, and Symmetrical Uncertainty.
CFS computes the correlation value existing between features and features with
their classes, as shown in Eq. (1). To select a set of features, these features and their
classes have higher correlation and lower correlation values between these features.
krcf
Merits ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ð1Þ
k þ kðk 1Þkrff
where k represents the number of features, rcf is the average correlation value between
features, and their class, rff shows the average correction value between features.
The Correlation method computes the value of Pearson’s correlation coefficient
between features and their classes. A higher value means that the feature is more
suitable to predict the class of input data. The ranking list of the feature set, including
the preset number of features, follows the descending order of correlation values.
The Information Gain method evaluates the feature worth as shown in Eq. (2).
A higher value also means that it is good for data classification. Generally, Information
Gain contains features with different values, but the values do not signify importance,
for example source IP or destination IP.
IGðClass; AttributeÞ ¼ HðClassÞ HðClassjAttributeÞ ð2Þ
where H stands for entropy value.

The Gain Ratio is adopted to evaluate the feature’s worth, as shown in Eq. (3).
Gain Ratio improves the problem of Information Gain. After normalization, Informa-
tion Gain reduces the deviation. A higher feature value means that the feature is better
to classify data.
H ðClassÞ HðClassjAttributeÞ
GRðClass; AttributeÞ ¼ ð3Þ
H ðAttributeÞ

The OneR [21] algorithm is applied to evaluate the feature worth by the error rate of
feature classification.
The ReliefF method [22] randomly selects samples from the dataset and considers
the features’ values, including the closest features belonging to the same class and
different classes to evaluate features’ weights, as shown in Eq. (4)
W½A diff ðA; R; HÞ X ½PðCÞ diff ðA; R; M ðCÞÞ

W½ A ¼ þ ð4Þ
m C6¼classðRÞ
m
where m is the number of samples, in diff(Attribute, instance1, instance2). The diff

function computes the difference between two instances. Instance2H is the closest
sample to instance1R in the same class. Instance2M is the closest sample to instance1
but belongs to a different class.
The Symmetrical Uncertainty method evaluates the correlation value between
features and their classes, as shown in Eq. (5)

H ð X Þ þ H ðY Þ H ðX; Y Þ
SU ¼ 2:0 ð5Þ
H ð X Þ þ H ðY Þ

782 Y.-J. Su et al.
The Wrapper method adopts learning algorithms to evaluate the feature dataset. In
the evaluation of k-folder cross-validation, each selected feature set has to be selected
by learning algorithms with results evaluated. Therefore, the process causes large
amounts of computation, but can select a feature set with higher quality.
UNSW-NB15 [17, 18] was adopted to be the experiment dataset in this study. The
main reason is that the network behavior of the dataset is similar to that of the current
network flow. The dataset includes nine kinds of the attack classifications, which is also
closer to current attack behaviors. The large amount of data in the dataset is expected to
lead to more reliable evaluation result.
The machine learning methods [23], once adopted, converted the intrusion detec-
tion dataset to meaningful data. First, seven machine learning methods (BayesNet,
NaiveBayes, JRip, PART, J48, RandomTree and RandomForest) of Weka were
adopted to train the attack models. Each method was applied to only part of the features
selected to construct feature patterns, which were in turn used to identify nine kinds of
network attacks.
The specification of experiment environment is CPU Intel(R) Core i7-7700,
memory 8 GB, operating system Microsoft Windows 10. The Weka version is 3.8. The
experiment dataset was extracted from UNSW-NB15 where 440,044 records were
randomly selected as the training dataset and 200,000 records as the testing dataset.
Seven machine learning methods of Weka were used to train attack patterns, but they
were applied only to part of the features selected by 4 feature selection methods, as
shown in Table 1. To train each attack pattern with higher accuracy, each machine
learning method adopted a different feature selection method to select a different
number of features, e.g. 9, 12, 15. In general, the number of features ranges from
9 to 15 [24, 25].
Table 1. Machine learning methods with feature selection

Machine Feature No. of Feature set
learning selection features
BayesNet ReliefF 12 1, 3, 4, 5, 10, 11, 14, 16, 23, 24, 37, 42
JRip Wrapper 15 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15
OneR Gain Ratio 9 1, 3, 6, 10, 11, 33, 34, 35, 37
PART Wrapper 9 1, 2, 3, 4, 5, 6, 7, 8, 9
J48 Wrapper 12 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12
RandomTree Correlation 9 6, 10, 19, 20, 24, 37, 42, 46, 47
RandomForest Correlation 15 5, 6, 10, 16, 19, 20, 21, 22, 24, 37, 41, 42, 45,
46, 47
Table 2 shows that seven machine learning methods were adopted with or without
feature selection methods to train the patterns of network attack behaviors. Most of the
time, the number of features from 9 to 15 had better results [25]. In the experiment, the
numbers of selected features were 9, 12, and 15. The measure results included recall,
precision, FNR (False Non-discovery Rate), AA (Average Accuracy), and MCC
(Matthews Correlation Coefficient). The execution time was effectively reduced when
using feature selection, and the value of AA also showed improvement.
FNR (False Non-discovery Rate): all anomalous network behaviors are erroneously
judged as normal network behaviors (the expect value is 0), as shown in Eq. (6).
FN
FNR ¼ 100 ð6Þ
TP þ FN
AA (Average Accuracy) can be computed, as shown in Eq. (7), and the expected
value of AA is 100.
TP þ TN
AA ¼ 100 ð7Þ
TP þ TN þ FP þ FN
The value of MCC (Matthews Correlation Coefficient) is distributed from −1 to 1,

as shown in Eq. (8). When the value is equal to 1, it means that the estimation is
correct, while 0 means the estimation is wrong.
TP TN FP FN
MCC ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ð8Þ
ðTP þ FPÞðTP þ FN ÞðTN þ FPÞðTN þ FN Þ
Table 2. Comparison of 7 machine methods

Machine Feature FNR AA Recall Precision MCC Execution time
learning selection % % % % (s)
BayesNet No 0.76 97.87 99.24 91.38 0.939 11.3
Yes 0.07 97.98 99.93 91.3 0.943 3.2
JRip No 2.41 98.68 97.59 96.16 0.96 9.4
Yes 2.83 98.88 97.17 97.48 0.966 2.9
OneR No 4 97.67 96 93.13 0.931 9.1
Yes 0.81 98.04 99.19 92.1 0.944 5.8
PART No 1.85 98.35 98.15 94.24 0.951 11
Yes 1.73 98.62 98.27 95.32 0.959 3.6
J48 No 0.57 98.33 99.43 93.09 0.952 8.7
Yes 2.43 98.72 97.57 96.4 0.962 3
RandomTree No 2.42 98.11 97.58 93.68 0.944 9.1
Yes 2.94 98.61 97.06 96.36 0.958 3
RandomForest No 0.92 98.41 99.08 93.72 0.954 46.1
Yes 2.85 98.62 97.15 96.31 0.959 10.8
784 Y.-J. Su et al.
5 Conclusion
The research proposed a real time IDS that comprises an offline phase and an online
phase, to detect malicious network behaviors with behavior patterns trained by 7
machine learning methods. To improve the performance and accuracy, 4 feature
selection methods were adopted in the proposed IDS. The experiment adopted the
UNSW-NB15 dataset, which includes 9 kinds of attack behaviors, to test the accuracy
of the IDS. Experimental results show that IDS using feature selection to train the
attack models can have better results in execution time and average accuracy.
References
1. Anderson, J.P.: Computer Security Threat Monitoring and Surveillance (1980)
2. Denning, D.E.: An intrusion-detection model. IEEE Trans. Softw. Eng. 13(2), 222–232
(1987)
3. Hall, M.A.: Correlation-based feature selection for machine learning. Doctoral Dissertation,
University of Waikato, Department of Computer Science (1999)
4. Mukkamala, S., Janoski, G., Sung, A.: Intrusion detection using neural networks and support
vector machines. In: Proceedings of the International Joint Conference on Neural Networks.
vol. 2, pp. 1702–1707 (2002)
5. Sung, A.H., Mukkamala, S.: Identifying important features for intrusion detection using
support vector machines and neural networks. In: Proceedings of International Symposium
on Applications and the Internet.vol. 217, pp. 209–216 (2003)
6. Kumar, K., Batth, J.S.: Network intrusion detection with feature selection techniques using
machine-learning algorithms. Int. J. Comput. Appl. 150(12), 1–13 (2016)
7. Amiri, F., Yousefi, M.R., Lucas, C., Shakery, A., Yazdani, N.: Mutual information-based
feature selection for intrusion detection systems. J. Netw. Comput. Appl. 34(4), 1184–1199
(2011)
8. Li, Y., Xia, J., Zhang, S., Yan, J., Ai, X., Dai, K.: An efficient intrusion detection system
based on support vector machines and gradually feature removal method. Expert Syst. Appl.
39(1), 424–430 (2012)
9. Ambusaidi, M.A., He, X., Nanda, P., Tan, Z.: Building an intrusion detection system using a
filter-based feature selection algorithm. IEEE Trans. Comput. 65(10), 2986–2998 (2016)
10. Scarfone, K., Mell, P.: Guide to intrusion detection and prevention systems (IDPS). NIST
Special Publication, 800-94 (2007)
11. Chandola, V., Banerjee, A., Kumar, V.: Anomaly detection: a survey. ACM Comput. Surv.
(CSUR), 41(3) (2009)
12. Depren, O., Topallar, M., Anarim, E., Ciliz, M.K.: An intelligent intrusion detection system
(IDS) for anomaly and misuse detection in computer networks. Expert Syst. Appl. 29(4),
713–722 (2005)
13. Kim, G., Lee, S., Kim, S.: A novel hybrid intrusion detection method integrating anomaly
detection with misuse detection. Expert Syst. Appl. 41(4), 1690–1700 (2014)
14. Cisco Visual Networking Index: Forecast and Trends, 2017–2022. https://www.cisco.com/c/
en/us/solutions/collateral/service-provider/visual-networking-index-vni/white-paper-c11-
741490.html#_Toc529314175
15. Blum, A.L., Langley, P.: Selection of relevant features and examples in machine learning.
Artif. Intell. 97(1–2), 245–271 (1997)
16. Moustafa, N., Slay, J.: UNSW-NB15: a comprehensive data set for network intrusion
detection systems (UNSW-NB15 network data set). In: Proceedings of Military Commu-
nications and Information Systems Conference (MilCIS), pp. 1–6 (2015)
17. Moustafa, N., Slay, J.: A hybrid feature selection for network intrusion detection systems:
Central points. In: Proceedings of the 16th Australian Information Warfare Conference,
pp. 5–13 (2015)
18. Moustafa, N., Slay, J.: The significant features of the UNSW-NB15 and the KDD99 data sets
for network intrusion detection systems. In: Proceedings of 4th International Workshop on
Building Analysis Datasets and Gathering Experience Returns for Security (BADGERS),
pp. 25–31 (2015)
19. Weka 3: Data Mining Software in Java. https://www.cs.waikato.ac.nz/ml/weka/
20. Hall, M.A., Holmes, G.: Benchmarking attribute selection techniques for discrete class data
mining. IEEE Trans. Knowl. Data Eng. 15(6), 1437–1447 (2003)
21. Holte, R.C.: Very simple classification rules perform well on most commonly used datasets.
Mach. Learn. 11(1), 63–90 (1993)
22. Kononenko I.: Estimating attributes: Analysis and extensions of RELIEF, Lecture Notes in
Computer Science, vol. 784, pp. 171–182. Springer, Heidelberg (1994)
23. Prachi, C.: Usage of machine learning for intrusion detection in a network. Int. J. Comput.
Netw. Appl. 3(6), 139–147 (2016)
24. Al-Jarrah, O.Y., Siddiqui, A., Elsalamouny, M., Yoo, P.D., Muhaidat, S., Kim, K.: Machine-
learning-based feature selection techniques for large-scale network intrusion detection. IEEE
34th International Conference on Distributed Computing Systems Workshops (ICDCSW),
pp. 177–181 (2014)
25. Rathore, M.M., Ahmad, A., Paul, A.: Real time intrusion detection system for ultra-high-
speed big data environments. J. Supercomput. 72(9), 3489–3510 (2016)
A CNN-Based Method for Infant Cry
Detection and Recognition
Chuan-Yu Chang(&) and Lung-Yu Tsai

National Yunlin University of Science and Technology, Douliu, Taiwan
{chuanyu,m10517035}@yuntech.edu.tw
Abstract. Crying is the primary means of communication between the baby

and the outside world. When a baby is crying, it is difficult for a novice parent to
immediately understand the baby’s needs. If parents can accurately determine
the cause of the baby’s cry, they can understand the baby’s emotional and
physiological changes and needs. In real-world applications, recording devices
may record sounds that are not produced by a baby. To reduce the burden on the
recognition server and improve the accuracy of the classifier, this study proposes
the conversion of the baby’s crying signal into a two-dimensional spectrogram.
A convolutional neural network is used to determine if the input spectrum
represents a baby’s cry. A baby’s cry is ultimately divided into four categories
(including pain, hunger, sleepiness, and wet diaper) through additional one-
dimensional convolutional neural networks. Experimental results showed that
the proposed method achieves high crying detection and recognition rates.
1 Introduction
In recent years, a lot of deep learning research had been applied to image recognition
and speech recognition. Researchers had also proposed many algorithms such as
convolutional neuron networks (CNN), Long Short Term Memory Network, LSTM,
GoogLeNet, and Generative Adversarial Network [1–4]. In addition, Yamamoto had
successfully identified infant’s cry using fast Fourier transform (FFT) and PCA [5–7].
Yizhar had great success on the recognition of infant cries through the combination of
algorithms such as the Mel Frequency Cepstral Coefficient (MFCC), Spectrogram and
CNN [8–11]. Based on the aforementioned research results, it is obvious that deep
learning is an effective and suitable algorithm for the discrimination of infant cries.
Therefore, in this work, the authors propose to use pre-emphasis, spectrogram, and
CNN to identify crying, as well as a two-stage method to classify the crying into four
categories.
2 Methodology
The proposed method differentiates between infant cries caused by pain, hunger,
sleepiness, and wet diapers using deep learning with two stages. Figure 1 shows the
system architecture of the proposed two-stage system. First of all, the sampling signal

https://doi.org/10.1007/978-3-030-15035-8_76
A CNN-Based Method for Infant Cry Detection and Recognition 787
S(n) is sequentially pre-emphasized, normalized and eventually resulting in S3 ðn). The

resulting S3 ðn) is sent to stage one to check if the signal represents an infant cry.
Whenever S3 ðn) is identified as an infant cry, it is sent to stage two to discriminate
between infant cries incurred by pain, hunger, wet diaper, or sleepiness. Otherwise, it is
discarded.
2.1 Pre-emphasis
The pre-emphasis stage refers to a system process designed to increase (within a
frequency band) the magnitude of some (usually higher) frequencies with respect to the
magnitude of other (usually lower) frequencies. Low frequency indicates slow varia-
tions in time and the numerical values of a low-frequency signal tend to change slowly
or smoothly from sample to sample. Due to the aforementioned property, Eq. (1) is
used to implement the pre-emphasis process (in this paper a ¼ 0:95) [12].
S2 ðnÞ ¼ SðnÞ a Sðn 1Þ ð1Þ
Figure 2 shows that the pre-emphasized signal becomes more pronounced in the
treble part, and the sound’s intervals become more defined.
788 C.-Y. Chang and L.-Y. Tsai
Fig. 2. Audio pre-emphasis results
2.2 Volume Normalization

The difference in distances between the infant and the equipment in recording time
could result in volume differences which affect the discriminability of the recording. To
improve the discriminability, volume normalization is applied to the signal. Volume
normalization is the application of a constant amount of gain to an audio recording to
bring the amplitude to a target level (the norm). Equation (2) is used to normalize the
amplitude of the signal to between –1 and 1.
S2 ðnÞ
S3 ðnÞ ¼ ð2Þ
maxðkmaxðS2 Þk; kminðS2 ÞkÞ
2.3 Stage 1: 2D CNN for Spectrogram Features

In this model, we use the audio spectrogram as input to the 2D CNN. A spectrogram is
generated by the STFT (Short Term Fourier Transform) of windowed audio or speech
signal. In this paper, the audio is sampled at 8000 Hz [13]. Each frame of audio is
windowed using a “tukey” window with a shape parameter of 0.25. The 2D CNN
architecture used to work with the spectrogram is presented in Fig. 3. The shape of the
spectrogram image entering the CNN is 65 156 and a softmax layer is used to
perform the two-class classification. In addition, “Adadelta” is used for training opti-
mization and the 2D CNN model applied to the spectrogram is presented in Fig. 4.
2.4 Stage 2: 1D CNN with Phoneme Input

In stage 2, down-sampling is performed on S3 ðn) using a lower rate of 1600 Hz. The
down-sampled signals are sent to a 1D CNN. Finally, the 1D CNN uses a softmax layer
to perform four-class classification. “Adadelta” is used for training optimization. The
1D CNN model working with S3 is presented in Fig. 5.
Fig. 3. Audio spectrogram
Fig. 4. Convolutional network model for infant cry detection.
Fig. 5. Convolutional network model for infant cry recognition.
2.5 Dataset
The infant cry dataset in this paper consists of samples recorded by first-time parents
using a smart phone with a mono channel, 8 kHz sampling rate and 16-bit resolution.
The samples were stored in WAV audio format. The dataset contains a total of 19691
samples. The length of each sound file is five seconds. The dataset was collected over
nine months from the birth of the infant. The dataset is tabulated in Table 1 and the
number in each field represents the samples recorded in a month.
Table 1. Crying file distribution table for different months

Months 0 1 2 3 4 5 6 7 8 9 Total
Pain 868 1143 916 872 395 284 243 133 56 535 5445
Hunger 832 1492 1265 886 541 309 191 147 87 513 6263
Sleepiness 679 986 871 791 434 270 182 151 68 495 4927
Wet diaper 191 226 117 2296 31 43 13 7 1 131 3056
Total 2570 3847 3169 4845 1401 906 629 438 212 1674 19691
3 Experiments
3.1 Infant-Cries Detection

The recorded sounds can be divided into three types: pure infant crying, infant crying
mixed with background noise and non-crying sound. Moreover, infant crying is sub-
divided into cries indicating Pain, Hunger, Sleepiness, and Wet diapers. In addition to
the infant crying, the recorded sound may also include background sounds such as TV,
speech, wind, music, and so on. The sounds are classified and shown in Table 2.
Table 2. Crying file distribution table for different months.

Infant crying Only infant crying • Pain
Infant crying mixed with background noise • Hunger
• Sleepiness
• Wet diaper
Non-infant crying • TV
• Speech
• Wind
• Music
• Fake crying
In order to maintain the balance of the training set during data training, 1500
samples randomly selected from each category of infant crying (Pain, Hunger,
Sleepiness, Wet diaper), were considered as positive crying samples. In addition,
another 1500 samples randomly selected from non-infant crying categories were
considered as negative samples.
3.2 Infant-Cries Recognition

In this paper, we used 5-fold cross-validation to divide the data into five equal parts.
The experimental results are shown in Table 3 [14]. Since infant crying changes with
age, the experiment was conducted independently every month. The experiment used
50% of the samples as the training set and 50% as the test set. The experimental results
are shown in Table 4.
Table 3. M2 K-fold cross-validation results.

K M2
1 77.25
2 77.22
3 78.39
4 78.51
5 79.99
AVG 78.27
Time (ms) 33.2961082458
Table 4. Infant crying demand accuracy analysis table.

Month 0 1 2 3 4
Prec. Recall Prec. Recall Prec. Recall Prec. Recall Prec. Recall
Pain 70 79 67 64 68 67 85 85 89 80
Hunger 71 69 64 78 72 72 83 83 78 90
Sleepiness 68 65 74 59 65 66 80 83 84 79
Wet diaper 83 61 93 61 75 70 100 99 100 72
Avg/Total 71 71 69 68 69 69 91 91 84 83
Month 5 6 7 8 9
Prec. Recall Prec. Recall Prec. Recall Prec. Recall Prec. Recall
Pain 79 56 70 59 74 68 67 52 65 68
Hunger 59 78 59 70 58 62 67 80 67 61
Sleepiness 70 65 60 65 59 62 76 71 62 68
Wet diaper 73 85 25 17 100 50 0 0 61 50
Avg/Total 69 67 63 63 64 63 70 69 65 65
4 Conclusion
Early stage infants can only express their emotions or physiological needs through
crying. Moreover, non-infant cry data will reduce the accuracy of infant-cry recogni-
tion. To solve the problem, this paper proposes a two-stage method to eliminate the
non-cry samples and recognize the cause of the cry. The first stage uses re-emphasis
and normalized mechanisms to convert the samples to spectrograms, which are used as
CNN model input. The proposed method achieves an accuracy of 99.83% for infant
crying detection. In the second stage, the second model divides the reasons for crying
into four types (Pain, Hunger, Sleepiness, Wet Diaper), and the recognition accuracy
reaches 78.27%.
References
1. Hochreiter, S., Schmidhuber, J.: Long short-term memory. Neural Comput. 9(8), 1735–1780
(1997)
2. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional
neural networks. In: Advances in Neural Information Processing Systems, pp. 1097–1105
(2012)
3. Ballester, P., de Araújo, R.M.: On the performance of GoogLeNet and AlexNet applied to
sketches. In: AAAI (2016)
4. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville,
A., Bengio, Y.: Generative adversarial nets. In: Advances in Neural Information Processing
Systems, pp. 2672–2680 (2014)
5. Rader, C., Brenner, N.: A new principle for fast Fourier transformation. IEEE Trans. Acoust.
Speech Signal Process. 24(3), 264–266 (1976)
6. Tyagi, V., Wellekens, C.: On desensitizing the Mel-Cepstrum to spurious spectral
components for robust speech recognition. In: Proceedings of the IEEE International
Conference on Acoustics, Speech, and Signal Processing, (ICASSP 2005) , vol. 1 (2005)
7. Garcia, J.O., Garcia, C.R.: Mel-frequency cepstrum coefficients extraction from infant cry
for classification of normal and pathological cry with feed-forward neural networks. Neural
Netw. 4, 3140–3145 (2003)
8. Petroni, M., et al.: Identification of pain from infant cry vocalizations using artificial neural
networks (ANNs). In: Applications and Science of Artificial Neural Networks, vol. 2492.
International Society for Optics and Photonics (1995)
9. Yong, B.F., Ting, H.N., Ng, K.H.: Baby cry recognition using deep neural networks. In:
World Congress on Medical Physics and Biomedical Engineering 2018. Springer, Singapore
(2019)
10. Abdel-Hamid, O., et al.: Convolutional neural networks for speech recognition. IEEE/ACM
Trans. Audio Speech, Lang. Process. 22(10), 1533–1545 (2014)
11. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional
neural networks. In: Advances in Neural Information Processing Systems (2012)
12. Muda, L., Begam, M., Elamvazuthi, I.: Voice recognition algorithms using Mel Frequency
Cepstral Coefficient (MFCC) and Dynamic Time Warping (DTW) techniques. arXiv preprint
arXiv:1003.4083 (2010)
13. Sohn, J., Sung, W.: A voice activity detector employing soft decision based noise spectrum
adaptation. Acoust. Speech Signal Process. 1, 365–368 (1998)
14. Fushiki, T.: Estimation of prediction error by using K-fold cross-validation. Stat. Comput. 21
(2), 137–146 (2011)
The Sensor Calibration and Growth
Parameters Monitoring for Phalaenopsis
Cultivation
Ding-Horng Chen1(&), Rong-Show Shen2, Tsai-Rong Chang1,

Pei-Shan Lin2, and Tzu-Ying Wang1
1
Southern Taiwan University of Science and Technology, Tainan, Taiwan
{chendh,trchang,zywang0607}@stust.edu.tw
2
Department of Horticulture, National Chiayi University, Chiayi, Taiwan
garden@mail.ncyu.edu.tw, s1000054@alumni.ncyu.edu.tw
Abstract. In this paper, a sequence of experiments that calibrate and collect the
sensor data that monitoring the growth of the orchids are derived. the sensor’s raw
data is calibrated and compared by using the popular and trusted WET2-2-K1
sensor. We also develop a cloud server and an analysis algorithm that collect the
data via different sensors by IoT technologies. A front-end responsive webpage
that visualize the Phalaenopsis physiological sensing parameters such as tem-
perature, moisture and conductivity of the cultivated soil are built, users can
monitor data on various devices such as personal computers, laptops or smart
hands to control and monitor Phalaenopsis’s growth environment data at any
time.
1 Introduction
According to the United Nations Commodity Trade Statistics Database (UN Comtrade),
orchids have a global output value of $1.7 billion in 2015 [1]. The Phalaenopsis flowers
have beautiful appearance, long flowering period and diverse colors, and have the
highest economic value in Taiwan. Taiwan is located in the subtropical zone; the good
weather conditions make Taiwan become an ideal growing environment for most
orchids. Taiwan is the northernmost boundary of the distribution of the original species
of Phalaenopsis. Over the past ten years, with the active guidance of the Taiwan’s
agricultural administrative units and the efforts of the orchid industry, the production
technology has increasingly improved. The transportation and marketing system have
gradually upgraded, and the planting area has also expanded year by year. The popu-
lation participating in cultivating orchids industry is the largest part in the flower
industry. According to the statistics of the Taiwan Customs Department (2018), the
export value of orchids in 2016 reached US$172 million, and Phalaenopsis accounts for
an export value of US$129 million.
The superior orchid planting process can ensure the quality, effectively reduce the
cost and increase the output value. It can not only promote the development of this
industry, but also increase the international competitiveness. Phalaenopsis has various

https://doi.org/10.1007/978-3-030-15035-8_77
794 D.-H. Chen et al.
growth-related factors from seedling to flowering stage, all the stages needs to be
properly controlled and analyzed to avoid pests or diseases to affect the flower quality.
In this paper, a series of reliable experiments were deduced that calibrate and
collect sensor data that monitors orchid growth. The raw data of the sensor is calibrated
and compared using the trusted and widely used Delta-T Devices WET2-2-K1 sensor
[2]. We use IoT technology to collect different sensor data and store it in the cloud
server. Users are allowed to control and monitor Phalaenopsis physiological sensing
parameters (such as temperature, moisture and soil conductivity) which are displayed
on the front-end responsive web page via various devices (PC, laptops or mobile
devices) at any time [3].
2 Background
The advantages of Taiwan’s orchid industry are mostly focused on the production
aspects, such as variety, climate and cultivation techniques. The most suitable tem-
perature for Phalaenopsis growth is between 25–30 °C. In summer, the temperature is
mainly controlled by the inner and outer shading nets, water walls and fans in the
greenhouse; while in winter, the heating machine is used to avoid the cold damage.
There is no complete and optimized physiological model can be intelligently referenced
and universally applied to other producers, places and varieties.
Taiwan’s production technology using sphagnum moss as a single cultivation
medium is quite sophisticated. However, the density, watering, frequency and nutrient
management methods of the sphagnum moss medium used by the industry are dif-
ferent, resulting in uneven seedling quality. Electrical conductivity (EC) is one of
the important factors affecting the vegetative growth of Phalaenopsis. The EC value is
the degree of conductivity of the salt dissociated in water, and can be regarded as the
concentration of soluble salts present in the medium. The higher the concentration, the
higher the osmotic potential; making crops difficult to absorb water and nutrients. As
the result, the root is damaged or leaves are withered. How to maximize the good
characteristics of the sphagnum moss medium, and to establish optimal physiological
parameters is important to improve the quality of the seedlings.
The concept of smart agriculture is to employ sensing components such as
biosensor, environmental sensors and image recognition sensors to integrate the IT and
IoT technologies to reduce the labor demand in farming and provide a more efficient
model to cultivate crops [4]. The Phalaenopsis flowers are popular crop and owns the
high-value in Taiwan for its beautiful flower shape. The advantages of Taiwan’s orchid
industry are mostly focused on production aspects, such as flower variety, climate and
cultivation techniques. Orchids are cultivated in greenhouse, the main factor that
affecting the growth condition contains the temperature, the moisture, the illumi-
nometer, soil moisture and the carbon dioxide density.
The first usage of IoT technology in the Phalaenopsis environmental monitoring
has adopted a non-destructive method to measure the growth parameters. In 2014, Peng
et al. [5] proposed a method to measure physiological parameters by using the spec-
trometer. The NTU team proposed a greenhouse monitoring system that applies the IoT
techniques such as wireless environment monitoring, image acquisition and back-end
The Sensor Calibration and Growth Parameters Monitoring for Phalaenopsis Cultivation 795
analysis system. Sensor data is collected and transmitted by ZigBee, an alternative to

short range and low power consumption data transmission. The sensor data is stored in
a back-end server, which is used to analyze the environmental factors of orchid cul-
tivation with statistical algorithm. Because of the large size of the greenhouse, many of
these transmitters need to be allocated and thus cause a high construction cost. This
system can only monitor the environmental factors in the greenhouse and cannot
monitor the physiological parameters of orchids in the sphagnum moss medium.
The Internet of Things (IoT) is a message exchange network that connecting dif-
ferent equipment such as sensors, mobile devices, appliances and servers [6]. The IoT
system collects, aggregates and analyzes data acquired from different data sources, can
be applied in many fields such as disaster prediction, disease control, artificial intel-
ligence, industry and agriculture, etc. The popular “Industry 4.0” concept wish to
construct a new intelligent industrial framework with perceptual awareness. By ana-
lyzing various huge amounts of data, it can generate a solution that fully satisfies
customers’ demand and reduce unnecessary cost. The IoT technique is an indispensable
component for Industry 4.0, and is gradually applied to agriculture. Engineers tried to
design a system to control the greenhouse environment and integrate data from various
plant physiological sensors. The system will control greenhouse environmental con-
ditions through programmable controllers (PLCs) and other device such as shading
nets, fans, water walls, fluids and irrigation equipment.
The current crop sensors are designed for general purpose, and there are no specific
sensors designed for Phalaenopsis on the market. The medium used for the cultivation
of Phalaenopsis is not the general soil but the sphagnum, so the general sensor cali-
bration standard is not fully applicable to Phalaenopsis. We will calibrate the sensor
accuracy required for cultivation of Phalaenopsis, such as soil temperature, soil
moisture, illuminometer and conductivity, so that the sensor can be used exclusively for
Phalaenopsis. The IoT data generated by these sensors will be transmitted to the cloud
server for big data integration and analysis so that researchers can control and monitor
the growth environment of Phalaenopsis at any time. The data will be visualized as a
chart; the trend of the environmental data can be seen at a glance.
3 Calibration of Phalaenopsis Sensors
In this study, we selected the WET2-2-K1 sensor (as shown in Fig. 1(a)) from Delta-T
Devices as the benchmark to calibrate the proposed sensors. The WET2-2-K1 is widely
used in horticulture and plant research, but this sensor lacks data upload function.
Therefore, when performing the calibration process, we must first manually measure
and record the measurement time, HubID and other data, and then upload it to the cloud
server. The newly developed moisture and conductivity sensor from Hugreen [7] are
employed for calibration purpose. Hugreen is a startup company that building precision
agriculture solutions to equip greenhouses and smallholder farmers with the tools
needed to become profitable. The proposed moisture and EC sensors are shown in
Fig. 1(b)(c). Since no sensor is perfect, the sensor must be calibrated before use for the
best accuracy. The signals of all measurement systems are affected by random noise or
signal lag. The reasons for calibration the sensors are to verify the precision (same input
should derive the same output) and resolution (reliably detect the small change in
measurement parameter) [8, 9]. In this study, we will perform a calibration of the
moisture and conductivity sensors. The calibration steps are as follows.
Fig. 1. The sensor used in this study: (a) WET2-2-K1 from Delta-T Devices (b) moisture sensor
from Hugreen (c) EC sensor from Hugreen
The Delta-T WET2-2-K1 has two measurement modes for measuring soil, namely
mineral and organic. Although the sphagnum moss is an organic material, the mineral
mode is often used in practical Phalaenopsis cultivation. In this study, we measure the
sensor data in two modes. The invalid samples of the sensors data (environmental tem-
perature, illuminometer, soil temperature, moisture and electrical conductivity) are first
removed, then we analyze the relation between the data from the Delta-T WET2-2-K1
sensor and the sensor to be calibrated.
The moisture sensor calibration method:

• Preparation: Sphagnum moss 200 ml, a 250 ml beaker
• The Delta-T WET2-2-K1 and the sensor to be calibrated were placed in the same
cultivation pot while measuring sensor moisture data for three different soil loca-
tions. Record the measurement time (year/month/day/hour: minute) and sensor
readings.
• The volume of sphagnum moss can be adjusted arbitrarily, but at least 50% of water
should be added. For example: 100 ml of sphagnum must be added to at least 50 g
of water, 200 ml of sphagnum must be added to at least 100 g of water, and so on.
• After adding water to the beaker, remove the sphagnum moss from the beaker. Rub
the moss so that the moss absorbs the water evenly and then put it into the beaker.
• Perform the following steps and measure the sensor data:
1. Place the Delta-T WET2-2-K1 and the sensor to be calibrated in the air, then
measure the sensor data.
2. Place the Delta-T WET2-2-K1 and the sensor to be calibrated in the beaker with
dry sphagnum moss in it, then measure the sensor data.
3. Place the Delta-T WET2-2-K1 and the sensor to be calibrated in the beaker with
5 g water, then measure the sensor data.
4. Repeat step 3 by adding 5 g water each time until at least 100 g water is added
(the theoretical moisture is 50%).
The moisture sensor calibration experiment result:
The experimental result of the moisture sensor calibration in mineral mode is
shown in Fig. 2. We can find that the correlation coefficient between Delta-T WET2-2-
K1 sensor and the proposed sensor is 0.77 before calibration; and the value become
0.86 after calibration. Although there is a difference between two sensors, the differ-
ence remains consistent; that means the accuracy of the proposed sensor is much
similar comparing with the Delta-T WET2-2-K1 sensor.
Fig. 2. The moisture sensor calibration result in mineral mode (a) before calibration (b) after
calibration
The experimental result of the moisture sensor calibration in organic mode is shown
in Fig. 3. The correlation coefficient is 0.75 before calibration; however, the correlation
coefficient is 0.19 due to an unexpectable error condition. The left part result in Fig. 3(b)
shows that the data of the two sensors are very consistent and the experiment design is
feasible. This result leaves a problem that has yet to be resolved.
Fig. 3. The moisture sensor calibration result in organic mode (a) before calibration (b) after
calibration
The electrical conductivity (EC) sensor calibration method:

• Preparation: eight 200 ml sphagnum moss, eight 250 ml beakers
• The Delta-T WET2-2-K1 and the sensor to be calibrated were placed in the same
beaker while measuring sensor EC value for three different locations. Record the
measurement time (year/month/day/hour: minute) and sensor readings.
• After adding diluted fertilizer solution (DFS) to the beaker, remove the sphagnum
moss from the beaker. Rub the moss so that the moss absorbs the diluted fertilizer
solution evenly and then put it into the beaker.
• Let
s p
V¼ ð1Þ
n
where s is the volume of sphagnum moss, p is the theoretical water percentage and
n is the number of the beakers. For example, if the volume of the sphagnum moss is
200 ml, the theoretical water percentage is 30%, the number of the beakers is 8, then
200 0:8
V¼ ¼ 7:5
8
1. Experiment 1: Add 60 g of the 30% water percentage DFS to the sphagnum moss to
the beaker. Gradually dilute the concentration of the solution by adding water (7.5 g
water each time) as shown in Fig. 4(a).
Fig. 4. Electrical conductivity (EC) sensor calibration method

2. Experiment 2: Add 80 g of the 40% water percentage DFS to the sphagnum moss to
the beaker. Gradually dilute the concentration of the solution by adding water (10 g
water each time) as shown in Fig. 4(b).
3. Experiment 3: Add 100 g of the 50% water percentage DFS to the sphagnum moss
to the beaker. Gradually dilute the concentration of the solution by adding water
(7.5 g water each time) as shown in Fig. 4(c).
The electrical conductivity (EC) experiment result:
As shown in Figs. 5 and 6, we perform the electrical conductivity experiment in
mineral mode and organic mode, respectively. The Ecb is the measured data and Ecs is
the converted data by Hugreen’s transformation equation. Comparing with the data by
Delta-T WET2-2-K1, the correlation coefficient before calibration is better than the
correlation coefficient after calibration. Possible causes of unsatisfactory results include
the poor design of the experimental method, errors in the measurement process, and
new sensor design issues. The results show that the proposed calibration method still
needs more experiments to verify its feasibility.
Fig. 5. The EC sensor calibration result in mineral mode (a) before calibration (b) after
calibration
Fig. 6. The EC sensor calibration result in organic mode (a) before calibration (b) after
calibration
4 Application of IoT Technology to Cultivation

of Phalaenopsis
The binary raw data (including soil temperature, moisture, illuminometer and con-
ductivity) returned by the Phalaenopsis sensor installed in the greenhouse (as shown in
Fig. 7) will be converted to decimal by the program. Users can periodically set up
sensors to capture data every 1 min, 10 min, 30 min or 1-h intervals. The sensor
specifications are shown in Table 1.
Fig. 7. The sensor configuration
Table 1. The sensor specifications

Environmental sensor Electrical Soil temperature and
conductivity sensor moisture sensor
Precision Illuminometer: ±5% ±0.05 ds/m Temperature: ±1 °C
Temperature: ±1 °C Moisture: ±3%
Moisture: ±3%
Display range Illuminometer: 1 Lux 0.01 ds/m Temperature: 1 °C
Temperature: 1 °C Moisture: 1%
Moisture: 1%
Working −20*75 °C
environment
The data collected by the sensor is transmitted from the hub to the gateway via
LoRa (Long Range) technology and then transmitted to the cloud database via Wifi.
LoRa is a low-power wide area network (LPWAN) communication protocol that
provides inexpensive remote connectivity for IoT devices in rural, remote and offshore
industries [10]. The overall architecture is illustrated in Fig. 8. In order to let users
query the growth history of Phalaenopsis, and to develop an artificial intelligence
algorithms in the future, we have set up a cloud server to store sensor data. The system
includes a back-end Apache web server, and the front-end web page accesses the
database via PHP. The sensor data is filtered according to the date selected by the user
and encoded into a JSON format. The application programming interface (API) is
provided to allow different front-end devices to display sensor data with a responsive
web page.
The sensor returns the environmental data to the cloud server, and user can access
the growth data of Phalaenopsis including user information, Hub number, Gateway
number, as well as power, environment temperature and humidity, illuminometer, soil
temperature and soil moisture of the sphagnum moss, and soil conductivity. The
growth parameters are shown in Fig. 9.
Fig. 8. The Phalaenopsis sensor architecture
Fig. 9. The growth parameters monitoring webpage

5 Conclusion
In this paper, a sequence of experiments that calibrate and collect the sensor data that
monitoring the growth of the orchids are derived. the sensor’s raw data is calibrated and
compared by using the popular and trusted WET2-2-K1 sensor. We also develop a
cloud server and an analysis algorithm that collect the data via different sensors by IoT
technologies. A front-end responsive webpage that visualize the Phalaenopsis physi-
ological sensing parameters such as temperature, moisture and conductivity of the
cultivated soil are built, users can monitor data on various devices such as personal
computers, laptops or smart hands to control and monitor Phalaenopsis’s growth
environment data at any time.
The experimental results show that the calibrated sensor data is indeed close to the
value of WET2-2-K1. However, the experimental results show that the newly devel-
oped sensor still has stability problems, which leads to sometimes the numerical dif-
ference is too large, which affects the correlation coefficient between the two sensors.
The experimental data obtained in this study will be provided to the sensor developer
for reference to further improve the stability of the sensor. At the same time, we will
experiment with different calibration methods and find other unexplained causes of
numerical fluctuations, so that the sensor can be used accurately for Phalaenopsis.
Acknowledgement. The author thanks the Ministry of Science and Technology in Taiwan for
funding (MOST 107-2321-B-415-005-) this study.
References
1. UN Comtrade (2015). http://comtrade.un.org/data/. Accessed 18 Sept 2015
2. WET KIT and WET-2 Sensor. https://www.delta-t.co.uk/product/wet-2-horticulture/
3. Gardner, B.S.: Responsive web design: enriching the user experience. Sigma J. Digit.
Ecosyst. 11(1), 13–19 (2011)
4. Dan, L., Xin, C., Chongwei, H., Liangliang, J.: Intelligent agriculture greenhouse
environment monitoring system based on IOT technology. In: 2015 International Conference
on Intelligent Transportation, Big Data and Smart City, pp. 487–490 (2015)
5. Peng, Y.-H., Hsu, C.-S., Huang, P.-C., Wu, Y.-D.: An effective wavelength utilization for
spectroscopic analysis on orchid chlorophyll measurement. In: 2014 IEEE International
Conference on Automation Science and Engineering (CASE), pp. 716–721 (2014)
6. Madakam, S., Ramaswamy, R., Tripathi, S.: Internet of Things (IoT): a literature review.
J. Comput. Commun. 3(05), 164 (2015)
7. Hugreen Co Ltd. https://hugreen.com.tw/
8. Sudduth, K.A., Drummond, S.T., Kitchen, N.R.: Accuracy issues in electromagnetic
induction sensing of soil electrical conductivity for precision agriculture. Comput. Electron.
Agric. 31(3), 239–264 (2001)
9. Quinones, H., Ruelle, P.: Operative calibration methodology of a TDR sensor for soil
moisture monitoring under irrigated crops. Subsurf. Sens. Technol. Appl. 2(1), 31–45 (2001)
10. LoRa. https://en.wikipedia.org/wiki/LoRa
The 4th International Workshop on Big
Data Processing in Online Social
Network (BOSON-2019)
Automatic Text Preprocessing
for Intelligent Dialog Agents
Alessandro Maisto1(B) , Serena Pelosi1 , Massimiliano Polito2 ,

and Michele Stingo1
1
Università di Salerno, Via Giovanni Paolo II, 84084 Fisciano, SA, Italy
{amaisto,spelosi,mstingo}@unisa.it
2
Network Contacts, Via Olivetti n◦ 17, 70056 Molfetta, BA, Italy
massimiliano.polito@network-contacts.it
Abstract. The paper describes a new Text Preprocessing Pipeline

based on a Hybrid approach which involve rule-based and stochastic
approaches. The presented pipeline is part of a larger project titled Big
Data for Multi-Agent Specialized System developed by Network Contacts
in collaboration with University of Salerno and other institutional part-
ners. The aim of the project is to build an Hybrid Question Answering
System composed by sets of Dialog Bots able to process great volumes of
data. Due to the importance of unstructured textual data, a particular
focus of the project is on automatic processing of Text. The paper will
describe the three main modules of the preprocessing pipeline, which
involve a Style Correction Module, a Clitic Decomposition Module and
a POS Tagging and Lemmatization Module.
1 Introduction
The present research focuses on a hybrid module for the preprocessing of Italian
texts. It represents a part of the research and the industrial project titled Big
Data for Multi-Agent Specialized System (BIG4MASS), which is developed by
the Network Contacts enterprise1 in collaboration with the Department of Polit-
ical and Communication Science of the University of Salerno and co-financed by
the Ministry of Economic Development. The BIG4MASS project concerns the
development of an Hybrid and distributed Question Answering System composed
by a set of Dialog Bots able to process great volumes of data and provided with
a Syntactic/Semantic Linguistic Engine.
In the last decade the information present on the web has increased more and
more. Talking about Big Data means talking about different kind of information
which can have a strong impact on different knowledge areas such as health [3],
government [1,2] or corporate. Due to the strategic importance of unstructured
data in information retrieval and domain specific knowledge management, a
1
Network Contacts is one of the major national player in BPO services (Business Pro-
cess Outsourcing), CRM (Customer Relationship Management), Digital Interaction
and Call & Contact Center.
https://doi.org/10.1007/978-3-030-15035-8_78
806 A. Maisto et al.
particular focus of the project is on automatic processing of texts. The linguistic

Core of the BIG4MASS project is the preprocessing module which consists of a
software package developed in java. This paper will describe all the steps which
are unavoidable in order to give to unstructured texts a preliminary and basic
structure for further and more complex analyses.
The modules that compose the preprocessing package are the followings:
• Style Correction Module: it is focused on the automatic correction of common

type errors. Therefore, the project involves different kinds of Corpora, many
of which composed by User Generated Contents and characterized by a great
number of spelling and style errors;
• Clitics decomposition: it focuses on the decomposition of the clitic pronouns
from the verbs in the Italian agglutinated forms. The isolation of Italian clitic
particles (ci, gli, la, le, lo, mi, si, ti, vi, etc.) and their combination (me-lo,
te-lo, etc.) is mandatory for a complete and correct POS Tagging and for the
localization and interpretation of the verbal arguments;
• Pos Tagging & Lemmatization: the POS Tagging is performed through TINT
[26], a parser already involved into the project. For what concerns the lemma-
tization phase it that takes advantage of the huge amount of linguistic data
provided by the Italian electronic linguistic resources developed in the Depart-
ment of Political and Communication Sciences of the University of Salerno.
The paper is structured as follows: Sect. 2 focuses on the literature related to

the three different modules presented in this paper; Sect. 3 describes the algo-
rithms we chose to deal with the problems posed by each module and the results
of the experimental phases.
2 State of the Art
In this Section we briefly mention the state of the art works related with the
three main modules of the preprocessing analysis.
Thanks to the increased availability of annotated corpora, it became possible
to train classifiers in order to automatically correct specific errors within texts
[16,29]. Classifiers models perform well over classes of single errors, failing when-
ever a classification system has to correct multiple (and often interconnected)
errors within a phrasal unit. To overcome this issue, hybrid methodologies have
been proposed [10]. In order to enhance sequential error analysis and to reach
better overall performances, Machine Translations algorithms have been tested
out. The logic behind automatic correction via machine translation is to con-
sider the correction itself as a translation task. Correction systems have two core
components: a linguistic model and a translation model [23]. Different kind of
models have been proposed such as hybrid approaches where a statistic algorithm
is integrated with a rule-based linguistic model [14], statistic approaches embed-
ded with the Levenshtein’s distance measure [19] and neural network approaches
[5,35,36].
Automatic Text Preprocessing for Intelligent Dialog Agents 807
As regards clitics, we define them as linguistic units that combine with other
words, called in literature “hosts”, with which they are phonologically related
[30]. In the present work, with “clitic” we mean just Italian clitic pronouns.
When clitics and verbs are agglutinated together they must be subjected to
a set of morpho-syntactic restrictions [30]. A software able to correctly analyze
and decompose clitic particles must take into account such constrains.
The present research focuses on agglutinated verbal forms, where verbs and
clitic particles appear to be tied together (with the clitic occurring after the
verb) into “monorematic” words (without any blank space between the words).
Such verbal forms, differentiated on the base of their moods, are of these kinds:
• infinitive mood (e.g. giocare + ci : giocarci, “to play with it/there”);

• gerund mood (e.g. giocando + ci : giocandoci “paying with it/there”);
• participle mood (e.g. giocato + ci : giocatoci “payed with it/there”);
• imperative mood (e.g. gioca + ci : giocaci “play with it/there”).
In all the other cases, which wont be object of the present research, clitics
and clitic clustering (sequences of more than one clitic disposed in succession),
occur before the verb, separated by blanks (e.g. ci si gioca).
Among the studies focused on clitics in Italian we mention [12,24,25,30,38].
As far as Pos Tagging is concerned, we must say that it is considered a
“solved task”, with state-of-the-art taggers achieving precision of 97%–98% [32,
37]. Among the best algorithms proposed in literature, we mention the Averaged
Perceptron methods [8], Stanford Tagger [22], SVMTool [15], BI-LSTM-CRF
[17], and NLP4J [6].
In Lemmatization task, tokens are converted into their common base form:
the lemma. In Italian language, the masculine singular is used as canonical form
of nouns, adjectives, pronouns and determiners, the present infinitive tense is
used for verbs. Often, Lemmatization can be replaced by Stemming, a procedure
that removes suffixes from a word form until the “stem” was found [28]. Actually,
both the approaches can be combined as in [18] for Icelandic language. Ripple
Down Rules [34] have been developed for knowledge acquisition and maintenance
of rule-based systems. The algorithm create exception to existing rules when a
rule produce incorrect conclusions. Most algorithms use external lexical informa-
tion, as Morfette [7], or SEM [9] for French texts. In particular, Morfette uses a
probabilistic system in order to joint morphological tagging and lemmatization
from morphologically annotated corpora. For what concern Italian resources, we
can cite an Italian version of TreeTagger [31], TagPro [27], CORISTagger [13],
Tanl POS tagger [4], ensemble-based taggers [11] and Pattern tagger [33].
3 Methodologies and Experiments
The present section will illustrate the methodologies involved in the development
of the three different modules of the project and the results of the experiments
carried out to proof the efficiency of the systems.
3.1 Style Correction
The system architecture exploits the EditDistance algorithm [20], which is

grounded on the knowledge provided by a dictionary of words written in the
correct way, which has been borrowed from the corpus Paisà.
The tool works into two different steps:
• StartUp: this phase is based on the distance algorithm and its purpose is to
prepare the data structure for the style correction. Starting from the correct
dictionary entry (e.g. acqua), the system checks the maxEditDistance value:
with the value of 1 the system cyclically subtracts each character of the word,
generating a list of possible errors (e.g. aqua; acua; cqua; acqu; acqa); with a
value of 2, the algorithm generates, in addition to the previously mentioned
list, another list where the items are created by cyclically subtracting couples
of characters from the correct word (e.g. aqu; cqu; aua; cua; acq, among
others). The purpose is to create a Knowledge Base (KB), created by the
tool, that includes an extended dictionary containing the lists of errors.
• Ready: this step includes two activities:
– the correction of simple words taken in isolation: a word W is compared
with the words contained in the correct dictionary. If there is a match with
a KB word, then the tools assumes that the word W has been written
in the proper way. Otherwise, it controls in the lists of errors contained
into the KB whereas there is a string close to W (DamerauLevenshtein
distance).
– correction of words connected by apostrophes: the system considers the
expressions composed by determiner/preposition + apostrophe + noun
just as simple words, to be listed and processed into the KB.
– the correction of multiword expressions (MWE)2 : the correction algo-
rithm treats these expressions just as patterns which are analyzed as a
whole, in order to avoid the alteration of the meaning of the MWE.
The Style Correction system has been tested on an experimental set of 1500
sentences (about 15,600 tokens) - each of which contained at least one or more
typing error - selected manually from an original corpus at disposal of Network
Contacts and made up of customer answers within the TEL-CO domain. As out-
lined in Table 1, the system corrected 2141 errors out of 2362. The vast majority
of these errors regarded the orthographic forms of single unit words, a small
amount were related to misspelled entity like smartphone names or telephone
promotion denomination which are usually lexically configured as multiword
expressions. The remaining errors can be explained as syntactical errors and
were not detected by the algorithm architecture.
2
With MWE we make reference both to lemmatized MWE, e.g. carta di credito, which
are listed in the Knowledge base, and to a set of MWE which are domain specific,
e.g. all inclusive unlimited, for the domain of telecommunications.
Table 1. Style correction’s results with differentiation between errors typologies
Error type Identified Unidentified

Orthographic 1931 43
Syntactic 0 31
Named entity 170 147
Total 2141 221
3.2 Clitic Decomposition
The method used to decompose agglutinated verb-clitic forms has been inspired
by the work of [38,39]. The pronominal particles under examination have been
classified, for computational purposes, on the base of the number of character
by which they are composed:
• 2 characters: e.g. lo, la, mi, vi, ti; mangialo “eat it”, prenditi “take it”;
• 3 characters: e.g. llo, cci, mmi, gli; fallo “do it”, dacci “give us”, digli “tell
him”;
• 4 characters: e.g. melo, mela, visi; scrivimelo “write it to me”, dettamela
“dictate it to me”, sedervisi “to sit down there”;
• 5 characters: e.g. ccelo, ttele, mmene; daccelo “give it to us”, fattele “do them
by yourself”, fammene “do them to me”;
• 6 characters: e.g. glielo, glieli, gliene; daglielo “give it to him/her”, scrivigliene
“write him/her about it”, mandaglieli “send them to him/her”.
Starting from this classification, the system recognizes all the clitic particles
as the result of the Cartesian Product of basic elements. This way we can rapidly
generate all the possible combination (from which are excluded all the errors).
As an example, given two lists of basic elements,
(a) me, te, se, ce, ve, mi, ti, ci, vi

(b) lo, la, li, le, ne, si, ci
it can be obtained the following combination:
a * b = [melo, mela, meli, mele, mene, mesi, meci, telo, tela, teli, tene, tesi,
teci, ...]
a * b represents the list of the particle composed by 4 characters. Some of

them are not grammatical, therefore are ignored a priori by the system.
The texts to be analyzed by the system are preprocessed in order to isolate
each word and verify the existence of clitics, starting from the longest sequences
(6 characters) and ending with the shorter ones (2 characters).
Given a candidate agglutinate verbal form, the system cuts off the particle
from each selected token. So, the initial part of the token is compared with the
verbal stems listed in the dictionaries; only if the system finds a match, therefore
the candidate is suitable to become a true positive agglutinate form verb+clitic.
The module aims at reducing, as much as possible, computational mistakes,
such as homography and false positivity. Examples are the following:
• “dalla” in “dalla a me, la penna”, “give it to me, the pen”, where the token is
a true agglutinated form of “dare+la”; and “dalla finestra alla porta”, “from
the window to the door”, where the token is just a preposition “da” combined
with the definite article “la”);
• “mangiatene” in “mangiate-ne”, where the token agglutinates a participle
“mangiate” and the particle “ne”; and “mangia-te-ne” in which the token
agglutinates the imperative “mangia” and the particles “te” and “ne”.
With the first example, a syntactic analysis performed on the sentence could
disambiguate the agglutinated forms, but in the second case it will be impossible
to detect when the word is an agglutinated form or not, also with a deep context
analysis performed.
The effectiveness of the method has been tested on 10,000 sentences (500,623
tokens) which have been randomly extracted from the corpus Paisà [21]. The
module for the clitic decomposition recognized and analyzed 5,544 verbs and
5,496 verbal forms agglutinated with clitics. Table 2 shows the evaluation of the
module under examination, which has been compared with the performances
of two of the most used Italian parsers. The module emerges in term of Recall,
while it presents a small deficiency with the Precision measure. This is due to the
fact that the module works on single tokens, before the postagging is performed,
without considering the context of tokens. This choice affects the analysis of false
positives like “influenzale” (that is recognized by our system as an agglutinated
form verb + clitic particle “influenza-le”, “influence them”, when is actually an
adjective “flu-like”, correctly postagged by the other parsers.
However, the Recall value reached by our tool makes it obtain the best F1
score (0.996) with respect to LinguA (0,982) e Tint (0,915).
Table 2. Clitics decomposition module results compared with LinguA and TINT
results
Tool TINT LinguA Clitics module

Precision 1 0,9978 0,9914
Recall 0,843 0,9665 1
F1 0,915 0,9819 0,9957
Accuracy 0,998 0,9996 0,9999
3.3 Lemmatization
The Lemmatization module is based on the DELAF, the Italian Electronic Dic-
tionary of Flexed Forms. It includes over 1 million of inflected Italian forms and
it has been divided into 6 sub-dictionaries in order to improve performance of
the algorithm.
In our methodology we directly match tokens with the tags that have been
inherited during the postagging process. Therefore, the system input is a list of
couples of token-postagging descriptions. Each token is matched into a dictionary
which is specific for the part of speech of interest in the list.
The information about the token that the lemmatizer will insert into the
database are of these kinds:
• inflectional and derivational morphology: e.g. +m for the masculine gender;

+ISSIMO for superlatives; +BILE for deverbal adjectives in −bile; +RI for
verbs reiterated with the prefix −ri, among others;
• Distributional classes for nouns: e.g. +Astr for abstract nouns; +Conc for
per concrete nouns; +Nedi for building structures o +Nloc for geographical
places, among others;
• other syntactic and semantic categories: e.g. +LOC for locative expres-
sions; +TEMP for the temporal ones; +t/+i for transitive/intransitive verbs;
+a/+e for the verb’s auxiliaries avere “to have” and essere “to be”.
Once that the token has been identified in the dictionary, comes the step of
the homonymy checking.
Dictionaries has been converted into a Json File in which each term is a
list. The first element of the list is the token, in order to easily extract all the
information once the POS Tagger analyze the word. The second element is a
Json Object in which all the morphosyntactical and semantic information about
the term is explicited. There are cases in which homonymy causes errors in the
lemmatization. However, it often happens that one of the results is obsolete (e.g.
sono that corresponds both to essere “to be” that sonare “to play”). In order to
quickly solve these problems we arranged a disambiguation list where we confer
a greater weight to the most common interpretations.
The Lemmatizator has been tested on a corpus of 250.000 tokens extracted
by a different portion of the “Paisá Corpus”. Due to the heterogeneous character
of the “Paisá Corpus” it is suitable for an efficient test phase.
To evaluate the module, we decided to compare its performances with Tree-
Tagger one of the most known tools for annotating text with part-of-speech and
lemma information for several languages.
We compared the tools evaluating the results obtained from the analysis on
the same data portion. An overview of the comparison is presented in Table 3.
POS Tag and Lemmatization module reaches very good results in terms of pre-
cision for lemmatization tasks. Conversely for what concerns the processing time,
the module still requires many adjustments. In fact, Tree-Tagger performs both
tasks concurrently and the processing time must be considered the same for POS
Table 3. Comparison of Lemmatization task
Tool Tree-Tagger Lemmatization module

Precision 0,927 0,969
Time 3,409 s 9,638 s
Tagging and Lemmatization. Otherwise, our module perform the two tasks subse-
quently, therefore the value must be added to the POS Tagging processing time.
At the end, we analyze the results in a qualitative way. The vast majority of
errors derives from an incorrect part-of-speech tagging in ambiguity cases (i.e.
condotta, V, lemmatized as condurre, “to lead”, instead of condotta, N that
could be lemmatized as condotta, N, “conduct”).
4 Conclusion
This research focused on the preprocessing step of a software which has been
developed for the linguistic interface of the BIG4MASS project. As this project
has been built around the Question Answering research topic, the linguistic
analysis phase is mandatory for the correct understanding of the questions and
the drafting of the answers. The preprocessing, first step in each text analysis,
could be crucial for the quality of the final results.
In our work we developed three main modules, which, starting from the style
correction, due to the characteristic of the domain corpora (in large part User
Generated Content), move thought the decomposition of agglutinated verb-clitic
forms, the POS Tagging and Lemmatization phase.
References
1. Amato, F., Colace, F., Greco, L., Moscato, V., Picariello, A.: Semantic processing
of multimedia data for e-government applications. J. Vis. Lang. Comput. 32, 35–41
(2016). https://www.scopus.com/inward/record.uri?eid=2-s2.0-84949036166&doi=
10.1016\%2fj.jvlc.2015.10.012&partnerID=40&md5=77d8fbfd4293af91d382e77055
3e159c.
2. Amato, F., Mazzeo, A., Penta, A., Picariello, A.: Knowledge representation and
management for e-government documents. In: IFIP International Federation for
Information Processing, vol. 280, pp. 31–40 (2008). https://www.scopus.com/
inward/record.uri?eid=2-s2.0-47249122655&doi=10.1007%2f978-0-387-09712-1
4&partnerID=40&md5=d8e43aeb328d1fdcb45aa323689f7010.
3. Amato, F., Moscato, F.: A model driven approach to data privacy verifica-
tion in e-health systems. Trans. Data Privacy 8(3), 273–296 (2015). https://
www.scopus.com/inward/record.uri?eid=2-s2.0-84949775424&partnerID=40&
md5=2ca02581d8482fbc2f8489a423a80290.
4. Attardi, G., Fuschetto, A., Tamberi, F., Simi, M., Vecchi, E.M.: Experiments in
tagger combination: arbitrating, guessing, correcting, suggesting. In: Proceedings
of Workshop Evalita, p. 10 (2009)
5. Bahdanau, D., Cho, K., Bengio, Y.: Neural machine translation by jointly learning
to align and translate. arXiv preprint arXiv:1409.0473 (2014)
6. Choi, J.D.: Dynamic feature induction: the last gist to the state-of-the-art. In:
Proceedings of NAACL-HLT, pp. 271–281 (2016)
7. Chrupala, G., Dinu, G., Van Genabith, J.: Learning morphology with Morfette
(2008)
8. Collins, M.: Discriminative training methods for hidden Markov models: theory
and experiments with perceptron algorithms. In: Proceedings of the ACL-02 Con-
ference on Empirical Methods in Natural Language Processing, vol. 10, pp. 1–8.
Association for Computational Linguistics (2002)
9. Constant, M., Tellier, I., Duchier, D., Dupont, Y., Sigogne, A., Billot, S.: Intégrer
des connaissances linguistiques dans un crf: application à l’apprentissage d’un
segmenteur-étiqueteur du français. In: TALN, vol. 1, p. 321 (2011)
10. Dahlmeier, D., Ng, H.T.: Better evaluation for grammatical error correction. In:
Proceedings of the 2012 Conference of the North American Chapter of the Associ-
ation for Computational Linguistics: Human Language Technologies, pp. 568–572.
11. Dell’Orletta, F.: Ensemble system for part-of-speech tagging. In: Proceedings of
EVALITA, vol. 9, pp. 1–8 (2009)
12. Delmonte, R.: Generating and parsing clitics with getarun. In: CLIN (1999)
13. Favretti, R.R., Tamburini, F., De Santis, C.: CORIS/CODIS: a corpus of written
Italian based on a defined and a dynamic model. A Rainbow of Corpora: Corpus
Linguistics and the Languages of the World. Munich: Lincom-Europa (2002)
14. Felice, M., Yuan, Z., Andersen, Ø.E., Yannakoudakis, H., Kochmar, E.: Grammat-
ical error correction using hybrid systems and type filtering. In: Proceedings of
the Eighteenth Conference on Computational Natural Language Learning: Shared
Task, pp. 15–24 (2014)
15. Giménez, J., Marquez, L.: Svmtool: a general POS tagger generator based on
support vector machines. In: Proceedings of the 4th International Conference on
Language Resources and Evaluation. Citeseer (2004)
16. Han, N.R., Chodorow, M., Leacock, C.: Detecting errors in English article usage
with a maximum entropy classifier trained on a large, diverse corpus. In: LREC
(2004)
17. Huang, Z., Xu, W., Yu, K.: Bidirectional LSTM-CRF models for sequence tagging.
arXiv preprint arXiv:1508.01991 (2015)
18. Ingason, A.K., Helgadóttir, S., Loftsson, H., Rögnvaldsson, E.: A mixed method
lemmatization algorithm using a hierarchy of linguistic identities (HOLI). In:
Advances in Natural Language Processing, pp. 205–216. Springer, Heidelberg
(2008)
19. Junczys-Dowmunt, M., Grundkiewicz, R.: The AMU system in the CoNLL-2014
shared task: grammatical error correction by data-intensive and feature-rich statis-
tical machine translation. In: Proceedings of the Eighteenth Conference on Com-
putational Natural Language Learning: Shared Task, pp. 25–33 (2014)
20. Levenshtein, V.I.: Binary codes capable of correcting deletions, insertions, and
reversals. Soviet Phys. doklady 10, 707–710 (1966)
21. Lyding, V., Stemle, E., Borghetti, C., Brunello, M., Castagnoli, S., Dell’Orletta,
F., Dittmann, H., Lenci, A., Pirrelli, V.: The PAISA corpus of Italian web texts.
In: Proceedings of the 9th Web as Corpus Workshop (WaC-9), pp. 36–43 (2014)
22. Manning, C.D.: Part-of-speech tagging from 97% to 100%: is it time for some
linguistics? In: International Conference on Intelligent Text Processing and Com-
putational Linguistics, pp. 171–189. Springer, Heidelberg (2011)
23. Mizumoto, T., Hayashibe, Y., Komachi, M., Nagata, M., Matsumoto, Y.: The
effect of learner corpus size in grammatical error correction of ESL writings. In:
Proceedings of COLING 2012: Posters, pp. 863–872 (2012)
24. Monachesi, P.: A grammar of Italian clitics (1996)
25. Monachesi, P.: A lexical approach to Italian cliticization (1999)
26. Palmero Aprosio, A., Moretti, G.: Italy goes to Stanford: a collection of CoreNLP
modules for Italian. ArXiv e-prints (2016)
27. Pianta, E., Zanoli, R.: TagPro: a system for Italian PoS tagging based on SVM.
Intelligenza Artificiale 4(2), 8–9 (2007)
28. Porter, M.F.: An algorithm for suffix stripping. Program 14(3), 130–137 (1980)
29. Rozovskaya, A., Roth, D.: Algorithm selection and model adaptation for ESL cor-
rection tasks. In: Proceedings of the 49th Annual Meeting of the Association for
Computational Linguistics: Human Language Technologies, vol. 1, pp. 924–933.
30. Russi, C.: Italian clitics: an empirical study, vol. 193. Walter de Gruyter (2008)
31. Schmid, H.: Treetagger—a language independent part-of-speech tagger. Institut
für Maschinelle Sprachverarbeitung, Universität Stuttgart 43, 28 (1995)
32. Shen, L., Satta, G., Joshi, A.: Guided learning for bidirectional sequence classifi-
cation. In: ACL, vol. 7, pp. 760–767. Citeseer (2007)
33. Smedt, T.D., Daelemans, W.: Pattern for python. J. Mach. Learn. Res. 13(Jun),
2063–2067 (2012)
34. Srinivasan, A., Compton, P., Malor, R., Edwards, G., Lazarus, L.: Knowledge
acquisition in context for a complex domain. Pre-print of Proceedings of the Fifth
EKAW91 (1991)
35. Sun, C., Jin, X., Lin, L., Zhao, Y., Wang, X.: Convolutional neural networks for
correcting English article errors. In: Natural Language Processing and Chinese
Computing, pp. 102–110. Springer, Heidelberg (2015)
36. Sutskever, I., Vinyals, O., Le, Q.V.: Sequence to sequence learning with neural
networks. In: Advances in Neural Information Processing Systems, pp. 3104–3112
(2014)
37. Toutanova, K., Klein, D., Manning, C.D., Singer, Y.: Feature-rich part-of-speech
tagging with a cyclic dependency network. In: Proceedings of the 2003 Conference
of the North American Chapter of the Association for Computational Linguistics on
Human Language Technology, vol. 1, pp. 173–180. Association for Computational
Linguistics (2003)
38. Vietri, S.: Dizionari elettronici e grammatiche a stati finiti: metodi di analisi for-
male della lingua italiana. Plectica (2008)
39. Vietri, S.: The Italian module for NooJ. In: Proceedings of the First Italian Con-
ference on Computational Linguistics, CLiC-it, pp. 389–393 (2014)
Exploiting Figures of Speech in Cultural
Heritage Reasoning
Flora Amato1(B) , Walter Balzano1 , Giovanni Cozzolino1 ,

and Francesco Moscato2
1
Department of Electrical Engineering and Information Technology,
University of Naples “Federico II”, Naples, Italy
{flora.amato,walter.balzano,giovanni.cozzolino}@unina.it
2
Department of Scienze Politiche, University of Campania “Luigi Vanvitelli”,
Caserta, Italy
francesco.moscato@unicampania.it
Abstract. Cultural Heritage are almost always commented and ana-

lyzed by using references to figures of speech. In particular, metaphors
and allegories are very frequent in ancient and historical documents,
painting and sculptures. It is frequent to have some hints about the
assets and their authors by comparing their contents with elements in
the domain of the figures of speech to which the asset refer. In order
to enable reasoning by figures of speech, we propose here a methodol-
ogy able to link concepts in the domain of cultural assets, with concepts
in the domain of the figure of speech. We show here how reasoning in
all these domains help in discover some elements related to the cultural
heritage that humans may neglect at first glance.
1 Introduction
Practically any asset of cultural heritage is drenched with metaphors and alle-
gories. For example, sovereigns have been represented or sculpted as mythologi-
cal figures. Gods where chosen to represent monarchs of various ages. Allegories
changed according to what was desired to be highlighted in the allegoric form:
Apollo-Sun was linked to sovereigns which wanted to emphasise the fact they
were the primary source of light for their kingdom and their subjects; Mars
exalted warrior and strategist properties of other monarchs (such as Napoleon).
Allegories in the iconography of sovereigns often were associated with their
lifestyle. Who was represented as the Sun God, maintained a more than lavish
life-style, who was painted as Mars was famous for his military victories. Still,
association to allegoric figures (always mythological ones) that were considered
negative.
Discovering hidden contents, authors provocations or clarifying unclear ele-
ments is easier when comparing elements in the asset domain with the metaphors
or allegories to which the elements are related. This is equivalent to reasoning in
abstract domain, and it is an old problem, but still investigated in literature [1,2].
https://doi.org/10.1007/978-3-030-15035-8_79
816 F. Amato et al.
One of the first attempt of coupling metaphors and analysis of assets is in [3].
Anyway the work outlined only the need of extending automated reasoning to
metaphors. The idea of using formal semantics (by ontologies) in metaphorical
reasoning is really novel [2,4,5]. Authors of [5] reported a model where metaphors
are associated to attributes in ontologies defined by OWL. The works in [2,4]
contains one of the first attempt of formalizing metaphors by using ontology-
based approaches.
However, at the best of our knowledge, previous works do not address the
problem of linking domains represented in cultural heritages with the ones
related to metaphors or allegories. The approach we present here is the first
attempt to overcome this problem.
2 Motivation and Related Works
For the Historical point of view, an attempt to apply trans-disciplinary case

studies as a means of sustainability of learning is presented in [6]. In this work
authors describe the theoretical importance of trans-disciplinary case studies in
learning historical science. The embryonic concepts of algorithm is hidden in
that methodology that evolves step-by-step, passing through a given number of
states. Trans-disciplinary definition of case studies, requests:
• An ontology conceptualizing of the phenomenon/problem/case to study or to

dealing with;
• An epistemology conceptualizing of the cognitive approaches the study is
relying on
• A methodology for the conceptualizing of the methods the study is applying
and their interaction with other methods
• A project management theory conceptualizing the approach to enact.
The same elements will be used while applying our approach, but it will
use ontology conceptualization of different knowledge domain in order to merge
ontologies by means metaphors. Metaphors identification will involve an con-
ceptualization of the cognitive approaches that lead to cope together symbols
from different knowledge domain. The methodology, which uses metaphors, will
involve interaction with formal semantics and semiotics domains.
It is clear that, in addition to the traditional tools for the study of social
science, Semantics (of documents) and Semiotics (a systemic relationship of signs
within communication processes) have to be used.
For historical and social sciences, the approach to the text must take into
account the relations between the languages and the speakers: words change their
meanings through the centuries and sometimes semantics related to phrases may
be lost. Researchers must take into account possible replacements of words and
meanings in order to re-build original semantics from historical documents.
The concept of meaning is important in linguistic theories. Each meaning is
defined only in relation to other meanings, within a given domain.
Exploiting Figures of Speech in Cultural Heritage Reasoning 817
In addition, even if an “object” changes the original name, it continues to

live linguistically. This leads to the need of paying particular attention to the
use and the origins of metaphors, metonymies and other figures of speech.
Languages and semantics of documents grown more and more during the
centuries in complexity. As described in [7], the semantics conceptualization of
relevant concepts was first represented on a tree data structure, but soon it
became insufficient to express concepts and a complex graph (assimilated to a
Labyrinth) takes its place.
In this tangle of semantic concepts and relations, we are trying to reduce
complexity by using Metaphors. In [8] the author reports an analysis of semiotic
and linguistic theories coming to the fundamental assertions about the limits
of semantic representations: they are too much complex to represent all the
meanings of all concepts in terms of languages elements which meanings changes
during centuries.
If the interpretation of the sources is demanded, in an hermeneutical perspec-
tive, to the symbols related to the words, instead of the meaning of the words,
complexity decreases and possibly new “lost” meanings of documents can be
retrieved.
3 Figures of Speech and Reasoning
The idea of matching both semantic and symbolic terminology for the descrip-
tion of the content of documents or archival artefacts opens new doors for what
regards their analysis. The allegoric figure, and generally that of the metaphor,
joins a meaning to a description when the description adopts terms and symbols
of a different domain to that of the intended meaning. If both the descriptive
and the metaphoric domain are described through ontologies, the first thing
that comes to mind is to associate, to concepts in the descriptive domain, those
of the metaphorical domain and, therefore, reason by analogy on the model in
the metaphorical domain, by adding some of the reasoning done to that of the
descriptive domain. It would be very useful to reason on documents by involv-
ing both domains. This way, connections could surface between concepts in the
descriptive domain that are difficult to understand if the reasoning was limited
to that domain, while are instead easy to highlight in the metaphorical domain.
Thanks to the use of formal (ontological) models [9–11] for the annotation of
documents, it is possible to introduce totally automatic reasoning processes,
namely create engines of artificial reasoning allowing to connect two domains of
interest and draw conclusions from their conjoined analysis.
One of the main reasons why an ontology is defined in OWL is to allow to
describe the content of a document or of an object in general, referring to some
sort of evolved vocabulary (that being the ontology), on which the most part of
users and scholars of such objects uniquely converge in the definition and use.
An Annotation is the link of ontologies elements to assets in cultural heritages
(whatever is the format used to store them: images, videos, digital documents
etc.) [12–15].
818 F. Amato et al.
For what regards annotations, there are different approaches. They can be
written directly into the document or through an external support (such as
another digital document); it may or may not allow the annotation through
individuals etc.
This is not the place to discuss all of the options, methodologies and tech-
niques used in the process of semantic annotation. A wide description can be
recovered from among the scientific literature.
What matters is that on the one hand domain experts are busy defining
ontologies to use within their scope of application; on the other, expert archivists,
historians, other humanities scholars, scientists etc. are working on connect-
ing the content of analysed documents to the concepts that are present within
domain ontologies. It is also clear that one document can be annotated for sev-
eral and different domain ontologies if the document is about concepts present
in more than one domain.
It is through annotation processes, and the analysis of the very annotations,
that it is possible to implement actions of semantic research within documents.
The procedures of semantic research are however difficult to apply to Web
documents, considering how many and how heterogeneous they are. They do
work well when related to narrow domains (or to a single semantic domain)
allowing for particular reasoning [16].
4 Metaphoric Abstraction
Metaphors are the basics elements of the methods we are goings to explain. The
definition of Metaphor, in the Oxford Dictionary, is the following: A figure of
speech in which a word or phrase is applied to an object or action to which it
is not literally applicable, or:A thing regarded as representative or symbolic of
something else.
The last definition is important because it contains the concept that a
Metaphor is something having a given semantics in a domain A, that is used to
address something else in an another domain B. The use, in the same document,
of usual concepts and metaphors, doubtless increases the complexity of the docu-
ment, but let us consider the case of a well-known, well-defined and clear domain
M, and a set of (historical) documents, from different periods or ages, belonging
to a knowledge domain D and that we suppose ‘metaphor-pruned’ (i.e. without
content containing metaphoric meanings). If we are able to describe semantics
of source documents in the domain D, in terms of the concepts of the domain M,
regardless of the meaning (i.e. semantics) of terms in the documents, we should
have the semantics of elements in the documents represented in terms of a clear
set of symbols (in his case words) with defined semantics [17–19]. Reasoning on
Domain M is simpler than reasoning on D and new concepts can be discovered
from source documents. More in detail, Abstraction, Pattern Recognition and
Generalisation are applied in our CT methodology. First of all, we have a set of
source documents (S) to analyse. Let us suppose that the documents belong to
different periods and that a domain ontology D exists able to describe the main
Fig. 1. Allegories elements
concepts in the documents. Let us assume that terms and concepts related to
them have changed during the years so D is no more able to describe all the
meanings of the documents since some of them have been lost due to language
changes (Fig. 1).
The first pass of the method is the retrieval of suitable Metaphors for doc-
uments contents. Let us characterise the Metaphors in ontological way and let
us assume that M is the ontology used for defining meanings of concepts in the
metaphoric domain. Both D and M have to be defined with a formal language.
It is extremely important that M is the more possible detailed. In our example
we use OWL in order to define ontologies. The next steps are the following: 1.
Documents are annotated by using concepts in D ontology. 2. The Abstraction
technique is applied: If a metaphor is known linking concepts (and instances)
from D to M, the link between concepts is formally reported. This like an oper-
ation of Ontology Merging, but D and M remain separated. At the end of this
step, it is possible to refer to known concept in D by using a metaphoric symbol.
3. The Pattern Recognition technique is then applied: If (graph) patterns exist
in M but there is no related pattern in D, even if similar relations (in terms
of graph structure) exist in the element of D metaphoric linked to elements in
M, probably the metaphoric connection has been lost during the years. 4. The
Generalisation technique is applied when the lost metaphoric link are re-applied
to source documents, recovering lost semantics [20–22] and information.
5 Conclusions and Future Works
In this paper we have defined a method for analysing by means of metaphors,

historical documents. Future works include the inclusion of automatic informa-
tion extraction techniques for (semi)-automatic annotation of source domain and
the joint use of different Metaphors on different ontologies in order to improve
obtained results.
820 F. Amato et al.
References
1. Narayanan, S.: Moving right along: a computational model of metaphoric reasoning
about events. In: AAAI/IAAI, pp. 121–127 (1999)
2. Diarra, D., Clouzot, M., Nicolle, C.: Causal reasoning and symbolic relationships
in medieval illuminations (2018)
3. Endert, A., North, C., Chang, R., Zhou, M.: Toward usable interactive analyt-
ics: coupling cognition and computation. In: Proceedings of the ACM SIGKDD
Workshop on Interactive Data Exploration and Analytics (IDEA 2014), pp. 52–56
(2014)
4. Gangemi, A., Alam, M., Presutti, V.: Amnestic forgery: an ontology of conceptual
metaphors. arXiv preprint arXiv:1805.12115 (2018)
5. Bulat, L., Clark, S., Shutova, E.: Modelling metaphor with attribute-based seman-
tics. In: Proceedings of the 15th Conference of the European Chapter of the Asso-
ciation for Computational Linguistics, Short Papers, vol. 2, pp. 523–528 (2017)
6. Posch, A., Scholz, R.W.: Transdisciplinary case studies for sustainability learning.
Int. J. Sustain. High. Educ. 7(3) (2006)
7. Eco, U.: Dall’albero al labirinto: studi storici sul segno e l’interpretazione. La Nave
di Teseo Editore spa, Milan (2017)
8. Eco, U.: Trattato di semiotica generale. La Nave di Teseo Editore spa, Milan (2016)
9. Amato, F., Moscato, F., Moscato, V., Colace, F.: Improving security in cloud by
formal modeling of IaaS resources. Future Gener. Comput. Syst. 87, 754–764 (2018)
10. Balzano, W., Murano, A., Vitale, F.: Hypaco–a new model for hybrid paths com-
pression of geodetic tracks. In: International Conference on Data Compression,
Communication, Processing and Security, CCPS-2016 (2016)
11. Cilardo, A.: Exploring the potential of threshold logic for cryptography-related
operations. IEEE Trans. Comput. 60(4), 452–462 (2011)
12. Piccialli, F., Chianese, A.: The internet of things supporting context-aware com-
puting: a cultural heritage case study. Mob. Netw. Appl. 22(2), 332–343 (2017)
13. Chianese, A., Marulli, F., Piccialli, F., Valente, I.: A novel challenge into mul-
timedia cultural heritage: an integrated approach to support cultural informa-
tion enrichment. In: 2013 International Conference on Signal-Image Technology &
Internet-Based Systems (SITIS), pp. 217–224. IEEE (2013)
14. Piccialli, F., Jung, J.E.: Understanding customer experience diffusion on social
networking services by big data analytics. Mob. Netw. Appl. 22(4), 605–612 (2017)
15. Amato, F., Moscato, V., Picariello, A., Piccialli, F.: SOS: a multimedia recom-
mender system for online social networks. Future Gener. Comput. Syst. (2017)
16. Mangold, C.: A survey and classification of semantic search approaches. Int. J.
Metadata Semant. Ontol. 2(1), 23–34 (2007)
17. Xhafa, F., Barolli, L.: Semantics, intelligent processing and services for big data.
Future Gener. Comput. Syst. 37, 201–202 (2014)
18. Moore, P., Xhafa, F., Barolli, L.: Semantic valence modeling: emotion recognition
and affective states in context-aware systems. In: 2014 28th International Confer-
ence on Advanced Information Networking and Applications Workshops (WAINA),
pp. 536–541. IEEE (2014)
19. Xhafa, F., Asimakopoulou, E., Bessis, N., Barolli, L., Takizawa, M.: An event-
based approach to supporting team coordination and decision making in disaster
management scenarios. In: 2011 Third International Conference on Intelligent Net-
working and Collaborative Systems (INCoS), pp. 741–745. IEEE (2011)
(2016)
21. Balzano, W., Vitale, F.: DiG-Park: a smart parking availability searching method
using V2V/V2I and DGP-class problem. In: 2017 31st International Conference
on Advanced Information Networking and Applications Workshops (WAINA), pp.
698–703. IEEE (2017)
22. Balzano, W., Murano, A., Stranieri, S.: Logic-based clustering approach for man-
agement and improvement of VANETs. J. High Speed Netw. 23(3), 225–236 (2017)
Textual Processing in Social Network
Analysis
Flora Amato1(B) , Walter Balzano1 , Giovanni Cozzolino1 , Alessandro de Luca1 ,

and Francesco Moscato2
1
Department of Electrical Engineering and Information Technology,
University of Naples “Federico II”, Naples, Italy
{flora.amato,walter.balzano,giovanni.cozzolino,
alessandro.deluca}@unina.it
2
Department of Scienze Politiche, University of Campania “Luigi Vanvitelli”,
Caserta, Italy
francesco.moscato@unicampania.it
Abstract. Social Networks are responsible of generating a huge amount

of information, intrinsically heterogeneous and coming from different
sources. In the social networks domain, the number of active users is
impressive, active users process and publish information in different for-
mats and data remain heterogeneous in their topics and in the published
media (text, video, images, audio, etc.). In this work, we present a general
framework for event detection in processing of heterogeneous data from
social networks. The framework we propose, implements some techniques
that users can exploit for malicious events detection on Twitter.
1 Introduction
The number of active users of social networking systems is amazing. Such kind
of systems can process and publish information with different formats. The pub-
lished information results inherently heterogeneous. Social media networks have
emerged as powerful means of communication for people looking to share and
exchange information on a wide variety of real-world events. These events range
from popular, widely known ones (e.g., a concert by a popular music band) to
smaller scale, local events (e.g., a local social gathering, a protest, or an acci-
dent). Short messages posted on social media sites such as Twitter can typically
reflect these events as they happen. For this reason, the content of such social
media sites is particularly useful for real-time identification of real-world events.
Social networks have become a virtually unlimited source of knowledge, that
can be used for scientific as well as commercial purposes. In fact, the analysis
of on-line public information collected by social networks is becoming increas-
ingly popular, not only for the detection of particular trend, close to the classic
market research problem, but also to solve problems of different nature, such
as the identification of fraudulent behaviour, optimization of web sites, tracking
the geographical location of particular users or, more generally, to find mean-
ingful patterns in a certain set of data. In particular, the rich knowledge that
https://doi.org/10.1007/978-3-030-15035-8_80
Textual Processing in Social Network Analysis 823
has accumulated in Twitter enables to catch the happening of real world events
in real-time. These event messages can provide a set of unique perspectives,
regardless of the event type [1–3], reflecting the points of view of users who are
interested or even participate in an event. In particular, for unplanned events
(e.g., the Iran election protests, earthquakes), Twitter users sometimes spread
news prior to the traditional news media [4]. Even for planned events, Twitter
users often post messages in anticipation of the event, which can lead to early
identification of interest in these events. Additionally, Twitter users often post
information on local, community-specific events where traditional news coverage
is low or nonexistent. Thus Twitter can be considered as a collector of real-time
information that could be used by public authorities as an additional informa-
tion source for obtaining warnings on event occurrence. In the last few years,
particular interest has given to the extraction and analysis of information from
social media by authorities and structures responsible for the protection of public
order and safety. Increasingly, through the information posted on social media,
public authorities are able to conduct a variety of activities, such as prevention
of terrorism and bio-terrorism, prevention of public order problems and safety
guarantee during demonstrations with large participation of people.
The heterogeneity of information and the huge scale of data makes the identi-
fication of events from Twitter a challenging problem. In fact, Twitter messages,
or tweets, has a variety of content types, including personal updates, not related
to any particular real-world event, information about event happenings, retweets
of messages which are of interest for a user and so on [5]. As an additional chal-
lenge, Twitter messages contain little textual information, having by design the
limit of 140 characters and often exhibit low quality [6]. Several research efforts
have focused on identifying events in Twitter [7,8]. For example, in recent years
there has been a lot of research efforts in analysing Tweets to enhance health
related alerts [9], following a natural calamity or a bio-terrorist attack, which
can urge a rapid response from health authority, as well as disease monitoring
for prevention. The previous work in this area includes validating the timeliness
of Twitter by correlating Tweets with the real-world statistics of disease activ-
ities, e.g., Influenza-like-Illness (ILI) rates [10–13], E-coli [14], cholera [15] or
officially notified cases of dengue [16]. In this work, we present a general frame-
work for event detection, starting from the heterogeneous processing of data
coming from social networks systems [17–20]. The proposed system process het-
erogeneous information [21–23] in order to detect anomalies in Twitter stream.
Event related anomalies are patterns in data that do not fit the pattern of the
expected normal behaviour [24]. Those anomalies might be induced in the data
for malicious activity, as for example cyber-intrusion or terrorist activity. The
proposed framework aims at analysing tweets and automatically extract relevant
event related information in order to raise a signal as soon as a malicious event
activity is detected. Ad-hoc hardware devices and high-performance communi-
cation scheduling can be adopted to manage big amount of data [25–28].
The reminder of the paper is structured as follows: in Sect. 2 we describe
two examples that motivate our work; in Sect. 3 we present an overview of the
824 F. Amato et al.
framework along with the goals and requirements that are of importance for
event detection from Twitter. Moreover in this Section we describe the main
processing stages of the system. In the end, in Sect. 4 we present some conclusions
and future work.
2 Motivating Example
Social Network, as Twitter, have become a virtually unlimited source of knowl-

edge, that can be used for scientific as well as commercial purposes. Twitter
users spreads a huge amount of information ranging on a several “facts”. The
analysis of the information collected by social networks is becoming increasingly
popular for the detection of trends, fraudulent behavior or, more generally, for
finding meaningful patterns in a certain set of data. This process of extraction
and analysis of large amounts of data is commonly defined data mining.
Particular attention has been paid to the analysis of the User-Generated Con-
tent (UGC) coming from Twitter, which is one of the most popular microblog-
ging websites. As described in the Sect. 1, Twitter textual data (i.e., tweets) can
be analyzed to discover user thoughts associated with specific events, as well
as aspects characterizing events according to user perception. Tweets are short,
user-generated, textual messages of at most 140 characters long and publicly
visible by default. For each tweet a list of additional features (e.g., GPS coor-
dinates, timestamp) on the context in which tweets have been posted is also
available.
In this Section we show two examples of text mining on live tweets to monitor
possible public security issues. In the first one we describe the topic relevance,
in the second one we analyze the results, using some realistic data taken during
a sport event.
In the last years, especially from the episodes following Sept. 11, the fear
of bioterrorism is in the list of most urgent healthcare concerns, together with
previous health issues - cancer, AIDS, smoking or alcohol/drug abuse, heart
disease and the cost of healthcare and insurance. Gallup1 has asked Americans
since 2001 to name the most urgent health problem facing the U.S. Cost of
and access to healthcare have generally topped the list in last few years, while
Americans most frequently mentioned diseases such as cancer and AIDS in the
1990s. Other health issues have appeared near the top of the list over the past
15 years, including 2014, when Ebola was listed among the top three health
concerns. This was most likely due to multiple Ebola outbreaks in West Africa
and a few confirmed cases all around the world, which prompted widespread
media coverage of the disease. In 2015, less than 0.5 of Americans listed Ebola
as the most urgent issue, as the threat of the virus has subsided. Similar short-
lived spikes in the responses to this question have occurred regarding other recent
health threats, including the H1N1/swine flu outbreak in 2009 and anthrax and
bioterrorism attacks in 2001.
1
http://www.gallup.com/.
In our first scenario, we consider as application domain a global health treat

like bioterrorism, so the system aim is to monitor the Twitter Stream detecting
tweets containing given keywords related to bioterrorism attacks, like “attack”,
“terror”, “anthrax, “terrorism”, “jihadi”, etc. In this scenario we focus our atten-
tion on gathering all the upcoming tweets about a particular event using Twit-
ter’ streaming API. Depending on the search term, we can gather tons of tweets
within a few minutes. This is especially true for live events with a world-wide
coverage (World Cups, Academy Awards, Election Day, you name it). A working
example that gathers all the new tweets with the #anthrax hashtag, or contain-
ing the word “anthrax”, shows that a large number of collected tweets is not
related to an upcoming real treat. In our case, for example, the word “anthrax”
is also related to a metal music band, so we need a technique to distinguish
relevant tweets from not relevant ones.
As for the second example, we consider a given event as search domain to
make a first selection of relevant tweets on which we adopt deeper textual analysis
to detect the positive ones. A post hoc analysis of the information posted on
Twitter, shows that, after the occurrence of a given event, the number of tweets
related to that event is influenced by two main factors:
1. density of users in different geographic areas;
2. popularity of the event in the different geographic areas.
To argue this consideration we consider, as a simplified example, a sport

event as given event; more specifically, we focus our attention to the soccer
match between SSC Napoli and Legia Varsavia teams, played on 10 December
2015 during qualifying stages of Europa League competition. Before the match,
tension, acts of vandalism and violence breaks out as opposite hooligans fight
in city centre. The news was reported by main national newspapers, as well as
main on-line information sites. Obviously this news was also reported by people
and mass media on their Twitter profiles.
We analyse relevant tweet concerning this given event, to estimate user per-
ception of the episodes of violence. Searching the string “Napoli Legia Varsavia”
and filtering tweets by date between 9 December and 11 December, we collect
the total amount of tweets concerning the given event. Refining the result with
the words “arresti”, “scontri”, “polizia”, etc., we obtain the number of tweets
related with the public security disorders, instead of the soccer match. In Table 1
we summary the results.
From this analysis we note that a considerable number of tweets is not related
with the event itself, but rather with the occurrence of acts of vandalism and
violence during the social event. Moreover, the proportion between total number
of tweets and the number of tweets concerning with public security issue depends
by the severity of the episodes.
826 F. Amato et al.
3 Framework Description
In this work we present a general framework to be adopted for event detection
from Twitter. In particular, we aim at detecting anomalies in Twitter stream
related to malicious activity, as for example cyber-intrusion or terrorist activ-
ity. In this section we present an overview of the main processing stages of the
Table 1. Search results.
Event related Public security concern

Date of interest: 9 December - 11 December
Query string: Napoli Legia Varsavia Napoli Legia Varsavia scontri,
Napoli Legia Varsavia arresti,
Napoli Legia Varsavia polizia
Number of tweets: 709 96
Fig. 1. Overview of the twitter stream processing framework for event detection and
alerting.
framework along with the goals and requirements that are of importance for
event detection from Twitter. Event detection from Twitter messages must effi-
ciently and accurately filter relevant information about events of specific interest,
which is hidden within a large amount of insignificant information. The pro-
posed framework aims at analyzing tweets and automatically extract relevant
event related information in order to raise a signal as soon as a malicious event
activity is detected. It is worth noting that in this work we consider an event
occurrence in Twitter stream whenever an anomaly bursts in the analyzed data
stream. The Twitter stream processing pipeline consists of three main stages,
as illustrated in Fig. 1. The system aims at detecting user defined phenomenon
form the twitter stream in order to give alerts in case of event occurrence. In the
following we describe the main stages of the proposed system.
3.1 Collection and Filtering

The first stage of the system is devoted to collect and filter tweets related to
an event. It is based on Information extraction task. Information extraction
is the process of automatically scanning text for information relevant to some
interest, including extracting entities, relations, and, most challenging, events
(something happened in particular place at particular time) [29]. It makes the
information in the text more accessible for further processing. The increasing
availability of on-line sources of information in the form of natural-language
texts increased accessibility of textual information. The overwhelming quantity
of available information has led to a strong interest in technology for processing
text automatically in order to extract task-relevant information. Information
extraction main task is to automatically extract structured information from
unstructured and/or semi-structured documents exploiting different kinds of text
analysis. Those are mostly related to techniques of Natural Language Processing
(NLP) and to cross-disciplinary perspectives including Statistical and Compu-
tational Linguistics [30–32], whose objective is to study and analyze natural
language and its functioning through computational tools and models [33–35].
Moreover techniques of information extraction can be associated with text min-
ing and semantic technologies activities in order to detect relevant concepts from
textual data aiming at detecting events, indexing and retrieval of information as
well as long term preservation issues. Standard approaches used for implement-
ing IE systems rely mostly on:
• Hand-written regular expressions. Hand-coded systems often rely on extensive

lists of people, organizations, locations, and other entity types.
• Machine Learning (ML) based Systems. Hand annotated corpus is costly thus
ML methods are used to automatically train an IE system to produce text
annotation. Those systems are mostly based on supervised techniques to learn
extraction patterns from plain or semi-structured texts. It is possible to dis-
tinguish two types of ML systems:
– Classifier based. A part of manually annotated corpus is used to train the
IE system in order to produce text annotation [36].
828 F. Amato et al.
– Active learning (or bootstrapping). In preparing for conventional super-

vised learning, one selects a corpus and annotates the entire corpus from
beginning to end. The idea of active learning involves having the system
select examples for the user to annotate which are likely to be informative
which are likely to improve the accuracy of the model [37]. Some examples
of IE systems are [38].
In our work we are interested in detecting specified event which relies on spe-
cific information and features that are known about the event such as type and
description, which are provided by the a domain expert of the event context.
These features are exploited by adapting traditional information extraction tech-
niques to the Twitter messages characteristics. In the first stage, the tweets are
collected exploiting the REST API provided from Twitter that allow program-
matic access to read data from the stream. Those API are costumed in order to
crawl data exploiting user defined rules (keywords, hashtag, user profile, etc.).
The tweets are annotated, in the filter module, with locations and temporal
expressions using a series of language processing tools for tokenization, part-
of-speech tagging, temporal expression extraction, and tools for named entity
recognition.
In order to access to Twitter data programmatically, it’s necessary to regis-
ter a trusted application, associated with an user account, that interacts with
Twitter APIs. Registering a trusted application, Twitter provides the required
credentials (consumer keys) that can be used to authenticate the REST calls via
the OAuth protocol. To create and manage REST calls we integrate in the Filter
Module a Python wrapper for the Twitter API, called Tweepy, that includes a
set of class enabling the interaction between the system and Twitter Stream.
For example, the API class provides access to the entire twitter RESTful API
methods. Each method can accept various parameters and return responses.
To gather all the upcoming tweets about a particular event, we call the
Streaming API extending Tweepy’s StreamListener() class in order to customise
the way we process the incoming data. In the Listing 1 we simply gather all the
new tweets with the #anthrax hashtag:
Listing 1. Custom Listener to gather all the new tweets with the #python hashtag
and save them in a file
class TwitterListener ( StreamListener ) :
def o n d a t a ( s e l f , data ) :
try :
with open ( ’ anthrax . j s o n ’ , ’ a ’ ) a s f :
f . w r i t e ( data )
return True
except BaseException as e :
print ( ” E r r o r o n d a t a : %s ” % s t r ( e ) )
return True
def o n e r r o r ( s e l f , s t a t u s ) :
print ( s t a t u s )
return True
t w i t t e r s t r e a m = Stream ( auth , T w i t t e r L i s t e n e r ( ) )
t w i t t e r s t r e a m . f i l t e r ( t r a c k =[ ’#anthrax ’ ] )
All gathered tweets will be furthermore filtered and then passed to the Clas-
sification module for deeper textual processing.
3.2 Classifying Relevant Information
The collected and annotated tweets are then classified, in the second stage, with
respect to their relevance. As stated before in the paper, event detection from
Twitter messages is a challenging task since the relevant information about a
specific event is hidden within a large amount of insignificant messages. Thus a
classifier must efficiently and accurately filter relevant information. The classifier
is trained exploiting a domain dependent thesaurus that include a list of relevant
terms for event detection. It is responsible for filtering out irrelevant messages.
3.3 Generating Alerts
Once the set of irrelevant tweets are discarded, those remaining must be con-
veniently aggregated with respect to an event of interest. The third stage of
the proposed system is aimed at detecting anomalies in the event related set
of tweets. Anomalies are patterns in data that do not fit the pattern of the
expected normal behavior. Anomalies might be induced in the data for a variety
of reasons, such as malicious activity, as for example cyber-intrusion or terror-
ist activity, but all of the reasons have a common characteristic that they are
interesting to the analyst. In literature there are several techniques exploited for
the anomaly detection. Those range between several disciplines and approaches
such as statistics, machine learning, data mining, information theory, spectral
theory.
In our proposed system, the set of event related tweets are processed by the
alert module that is responsible to raise an alert in case of anomaly within the
collected tweets. The alert module implements a set of burst detection algorithms
in order to generate alerts whenever an event occurs that means an anomaly in
the analyzed data stream is detected. The detected events are then displayed to
the end user exploiting the visualization module. An important aspect is that
the analyzed data can be tuned by the end user for novelty detection which aims
at detecting previously unobserved (emergent, novel) patterns in the data, such
as a new topic. The novel patterns are novel patterns are typically incorporated
into the normal model after being detected. Moreover with the tuning phase
the parameters for the tweets classification module can be modified in case the
messages that raised an alert are considered not relevant for the event occurrence.
830 F. Amato et al.

Social networking services are matter of interest for research activities in several
fields and applications. The number of active users of social networking services
like Twitter raised up to 320 million per month in 2015. The rich knowledge that
has accumulated in the social sites enables to catch the reflection of real world
events.
In this work we proposed a general framework for event detection, start-
ing from the processing of heterogeneous data coming from social networks sys-
tems. The framework aims at detecting malicious events in Twitter communities,
exploiting techniques related to heterogeneous data analysis and information
classification. Once the gathered data are classified as related to criminal activ-
ity, they are processed by the alert module that is responsible to raise an alert in
case of anomaly within the collected information. The alert module implements
a set of burst detection algorithms in order to generate alerts whenever an event
occurs that means an anomaly in the analyzed data stream is detected.
As future work we intend to provide the implementation of our system tuned
with a domain related to the bio-terrorism.
References
1. Diakopoulos, N., Naaman, M., Kivran-Swaine, F.: Diamonds in the rough: social
media visual analytics for journalistic inquiry. In: 2010 IEEE Symposium on Visual
Analytics Science and Technology (VAST), pp. 115–122, October 2010
2. Yardi, S., Boyd, D.: Tweeting from the town square: measuring geographic local
networks. In: International Conference on Weblogs and Social Media. American
Association for Artificial Intelligence, May 2010
3. Amato, F., Cozzolino, G., Mazzeo, A., Moscato, F.: Detect and correlate informa-
tion system events through verbose logging messages analysis. Computing (2018).
Cited By 0; Article in Press
4. Sakaki, T., Okazaki, M., Matsuo, Y.: Earthquake shakes Twitter users: real-time
event detection by social sensors. In: Proceedings of the 19th International Con-
ference on World Wide Web, WWW 2010, pp. 851–860. ACM, New York (2010)
5. Naaman, M., Boase, J., Lai, C.-H.: Is it really about me?: Message content in social
awareness streams. In: Proceedings of the 2010 ACM Conference on Computer
Supported Cooperative Work, CSCW 2010, pp. 189–192. ACM, New York (2010)
6. Becker, H., Naaman, M., Gravano, L.: Beyond trending topics: real-world event
identification on Twitter (2011)
7. Becker, H., Naaman, M., Gravano, L.: Learning similarity metrics for event iden-
tification in social media. In: Proceedings of the Third ACM International Con-
ference on Web Search and Data Mining, WSDM 2010, pp. 291–300. ACM, New
York (2010)
8. Atefeh, F., Khreich, W.: A survey of techniques for event detection in Twitter.
Comput. Intell. 31(1), 132–164 (2015)
9. Essmaeel, K., Gallo, L., Damiani, E., De Pietro, G., Dipanda, A.: Comparative
evaluation of methods for filtering kinect depth data. Multimedia Tools Appl.
74(17), 7331–7354 (2015)
10. Aramaki, E., Maskawa, S., Morita, M.: Twitter catches the flu: detecting influenza
epidemics using Twitter. In: Proceedings of Conference on Empirical Methods in
Natural Language Processing (2011)
11. Lamb, A., Paul, M.J., Dredze, M.: Separating fact from fear: tracking flu infections
on Twitter. In: HLT-NAACL, pp. 789–795 (2013)
12. Chon, J., Raymond, R., Wang, H., Wang, F.: Modeling flu trends with real-time
geo-tagged Twitter data streams. In: Wireless Algorithms, Systems, and Applica-
tions, pp. 60–69. Springer, Heidelberg (2015)
13. Amato, F., Cozzolino, G., Mazzeo, A., Romano, S.: Intelligent medical record man-
agement: a diagnosis support system. Int. J. High Performance Comput. Netw.
12(4), 391–399 (2018)
14. Diaz-Aviles, E., Stewart, A.: Tracking Twitter for epidemic intelligence: case study:
Ehec/hus outbreak in Germany. In: Proceedings of the 4th Annual ACM Web
Science Conference, WebSci 2012, pp. 82–85. ACM, New York (2012)
15. Chunara, R., Andrews, J.R., Brownstein, J.S.: Social and news media enable esti-
mation of epidemiological patterns early in the 2010 Haitian cholera outbreak. Am.
J. Trop. Med. Hyg. 86(1), 39–45 (2012)
16. Gomide, J., Veloso, A., Meira Jr, W., Almeida, V., Benevenuto, F., Ferraz, F.,
Teixeira, M.: Dengue surveillance based on a computational model of spatio-
temporal locality of Twitter. In: Proceedings of ACM WebSci 2011 (2011)
17. Piccialli, F., Chianese, A.: The internet of things supporting context-aware com-
puting: a cultural heritage case study. Mob. Netw. Appl. 22(2), 332–343 (2017)
18. Chianese, A., Marulli, F., Piccialli, F., Valente, I.: A novel challenge into mul-
timedia cultural heritage: an integrated approach to support cultural informa-
tion enrichment. In: 2013 International Conference on Signal-Image Technology &
Internet-Based Systems (SITIS), pp. 217–224. IEEE (2013)
19. Piccialli, F., Jung, J.E.: Understanding customer experience diffusion on social
networking services by big data analytics. Mob. Netw. Appl. 22(4), 605–612 (2017)
20. Amato, F., Moscato, V., Picariello, A., Piccialli, F.: SOS: a multimedia recom-
mender system for online social networks. Future Gen. Comput. Syst. (2017)
21. Xhafa, F., Asimakopoulou, E., Bessis, N., Barolli, L., Takizawa, M.: An event-
based approach to supporting team coordination and decision making in disaster
management scenarios. In: 2011 Third International Conference on Intelligent Net-
working and Collaborative Systems (INCoS), pp. 741–745. IEEE (2011)
22. Xhafa, F., Barolli, L.: Semantics, intelligent processing and services for big data
(2014)
23. Moore, P., Xhafa, F., Barolli, L.: Semantic valence modeling: emotion recognition
and affective states in context-aware systems. In: 2014 28th International Confer-
ence on Advanced Information Networking and Applications Workshops (WAINA),
pp. 536–541. IEEE (2014)
24. Chandola, V., Banerjee, A., Kumar, V.: Anomaly detection: a survey. ACM Com-
put. Surv. 41(3), 15:1–15:58 (2009)
26. Cilardo, A., Fusella, E., Gallo, L., Mazzeo, A.: Joint communication scheduling and
interconnect synthesis for FPGA-based many-core systems. In: Design, Automa-
tion and Test in Europe Conference and Exhibition (DATE), pp. 1–4. IEEE (2014)
27. Cilardo, A., Fusella, E., Gallo, L., Mazzeo, A.: Exploiting concurrency for the
automated synthesis of MPSoC interconnects. ACM Trans. Embedded Comput.
Syst. (TECS) 14(3), 57 (2015)
832 F. Amato et al.
28. Cilardo, A., Durante, P., Lofiego, C., Mazzeo, A.: Early prediction of hardware
complexity in HLL-to-HDL translation. In: 2010 International Conference on Field
Programmable Logic and Applications (FPL), pp. 483–488. IEEE (2010)
29. Hobbs, J.R., Riloff, E.: Information extraction. In: Indurkhya, N., Damerau, F.J.
(eds.) Handbook of Natural Language Processing, 2nd edn. CRC Press, Taylor and
Francis Group, Boca Raton (2010)
30. Butler, C.S.: Statistics in Linguistics. Blackwell, Oxford (1985)
31. Biber, D., Reppen, R., Conrad, S.: Corpus Linguistics: Investigating Language
Structure and Use. Cambridge University Press, Cambridge (1998)
32. Kennedy, G.D.: An Introduction to Corpus Linguistics. Longman, London (1998)
pression of geodetic tracks. In: The International Conference on Data Compression,
Communication, Processing and Security, CCPS 2016 (2016)
35. Balzano, W., Murano, A., Vitale, F.: SNOT-WiFi: sensor network-optimized train-
ing for wireless fingerprinting. J. High Speed Netw. 24(1), 79–87 (2018)
36. Riloff, E.: Automatically constructing a dictionary for information extraction tasks.
In: Proceedings of the Eleventh National Conference on Artificial Intelligence
(1993)
37. Grishman, R.: Information extraction: Capabilities and challenges (2012)
38. Pantel, P., Pennacchiotti, M.: Espresso: leveraging generic patterns for automati-
cally harvesting semantic relations. In: Proceedings of the 21st International Con-
ference on Computational Linguistics and the 44th Annual Meeting of the Associ-
ation for Computational Linguistics (2006)
ACOp: An Algorithm Based on Ant
Colony Optimization for Parking Slot
Detection
Walter Balzano(B) and Silvia Stranieri
Naples University, Federico II, Naples, Italy

walter.balzano@gmail.com, silviastranieri1047@gmail.com
Abstract. Ant Colony Optimization (ACO) is a known, largely

employed paradigm for optimization algorithm. It is a bio-inspired app-
roach following the real world ants beahvior in food search, that is trav-
eling along the path having the highest pheromone. ACO algorithms
provide a heuristic techinque to find global optima for an optimization
problem, by using the global parameters of trail and pheromone that
make a path more attractive than another. In this work, we focus on a
problem that has serius impact on traffic congestion in VANETs: avail-
able parking slot detection. Despite ACO paradigm has been largely used
in VANET field to address clusterization, routing, and communication
failure, parking problem has never been handled with ant colony opti-
mization. The main contribution of this paper is an innovative approach
to the parking detection, that is formulated as an optimization problem
and managed through ACO, and that provides, by means of opportune
representation of the environment, a path that maximizes the number of
available parking slot met.
Keywords: ACO · VANET · Optimization
1 Introduction
VANET has become an active research area due to its huge potential to improve
vehicle and road safety [4], and traffic control [6,27]. In such a network, vehicles
act as nodes of a graph where connections represent communication between
nodes: if two vehicle are able to communicate (i.e. they are in the respective
communication ranges), there is a connection between the corresponding nodes
[13]. Each vehicle can exchange network information with its direct neighbors,
while through multi-hop communication further vehicles can be reached [11,12].
Moreover, information exchange may be performed not only between vehicles,
but also between vehicles and some road infrastructure items, such as a traffic
light: indeed, typically two kind of communications are possible in a VANET:
Vehicle-to-Vehicle and Vehicle-to-Infrastructure [16,19].
With the purpose of improving traffic control in a VANET, in this work we
focus on parking detection. Indeed, the presence of wandering vehicles looking
https://doi.org/10.1007/978-3-030-15035-8_81
834 W. Balzano and S. Stranieri
for a free parking slot can significantly impact the traffic congestion in a VANET
and, beyond that, it constitutes a waste of time for drivers.
Automatic parking systems are made of different components, such as (i )
environment recognition, (ii ) path generation, and (iii ) path tracking [14,22].
In this work, we focus first on environment recognition phase, by providing a
structural representation of the considered area through a graph representing
the environment by regions of parking slot, and then on generation phase, by
applying ant colony optimization (ACO) to plan the optimal path leading to a
free parking slot. Our aim is to exploit a modeling approach, as in [5], whose
semantic [3] can be exploited to facilitate parking detection.
ACO is a paradigm for optimization algorithms that combines a priori knowl-
edge about promising solutions with a posteriori information of good solutions
[25]. This paradigm allows to build a complete, global, optimal solution through
successive approximations of good, local solutions.
The approach is bio-inspired, since it follows the behavior of ants in real
world: when ants look for food, they travel along the path having the highest
pheromone, which is the hormone secreted by other ants that have previously
taken that path [24].
Similarly, algorithms following this strategy are based on a loop performing
at each step opportune updates of global values in order to make a trail more
attractive than another one, according previous iterations results.
The ant colony optimization approach has been used already in literature to
address VANET challenges, such as clustering [9], and routing issues, but it has
never been applied for parking optimization purposes. The main contribution
of this work is a novel approach to the parking slot detection problem which
experiences a totally new combination with ant colony optimization.
Outline
The rest of the paper is organized as follows: in Sect. 2, an overview of existing
work is presented; in Sect. 3, we provide our ACOp approach to find the optimal
path with respect to the purpose of finding available parking slot; finally, in
Sect. 4, the conclusions are shown.
2 Related Work
Free parking slot detection is an interesting research field in VANETs context
[10,17], since it determines a big waste of time for drivers, and consequently
a traffic congestion increasing. For this reason, many proposals of automatic
parking systems are made by researchers trying to address the main VANET
issues, such as transmission delay, routing failure, and traffic congestion.
A lot of effort has been done by researcher to optimize network communica-
tion [7,8]. In [23], they point out that methods for target position-designation,
which is one of the core technologies in automatic parking systems, can be
divided into four categories: (i ) user-interface based, (ii ) parking slot mark-
ings based, (iii ) free-space based, and (iv ) infrastructure based. In [26], they
ACOp: An Algorithm Based on Ant Colony Optimization 835
propose a technique for vacant parking slot detection based on three stages: (i )
parking slot marking detection, (ii ) occupancy classification, and (iii ) parking
slot marking tracking. The first step recognizes different parking slot markings,
the second step identifies if the slots detected in the previous step are free, and
the third one estimates the position of the selected parking slot. Similarly, in
[22], they also provide an algorithm based on parking-slot-markings recognition.
They build a priori knowledge through one-dimensional filter in Hough space.
In our previous work [15], we analyze the distance geometry problem (DGP) and
we adopt it to the parking issue, by dividing vehicles in three groups according
to the possible states they can assume (parked, finding, and leaving), and man-
aging the interaction between these states in a probabilistic way through Markov
chains.
Ant Colony Optimization (ACO) is a paradigm for solving optimization prob-
lem that combines a priori information about promising solutions with a pos-
teriori information about previously obtained good solutions [25]. In literature,
we find different proposals applying ACO approach to VANET: in [1], they high-
lights that a VANET is more stable when an optimal number of clusters is built,
because network resources are efficiently employed, and for this reason they use
ant colony optimization applied to clustering techniques for VANETs. Results
show that ACO improves clustering performances, providing a technique which
is effective in comparison with others. Similarly, in [18], ACO is used to improve
VANET potentialities. In particular, their goal is to avoid the delay of com-
munication that occurs due to disconnection: for this reason, they apply ant
colony optimization to introduce repair algorithm in case of failure. In [24], they
use ACO to discover new routes in a VANET in an efficient way, with positive
results from performances point of view. In [21], they provide MACO, a modified
version of ant colony optimization, in which repulsion effect is used instead of
attractiveness in order to address the traffic congestion. In particular, they apply
ACO only in case of congestion, allowing vehicles to follow the non congested
path.
Despite the numerous application of ACO to vehicular context, to the best
of our knowledge, this is the first work addressing the free parking slot detection
by using bio-inspired ant colony optimization approach.
3 ACO for Parking
In this section, we want to propose a novel way to apply a family of solutions, the
Ant Colony Optimization, focusing on vehicular context, in particular parking
slot detection. Since free slot search constitutes a big waste of time for drivers,
many proposals to achieve good solutions to the problem have been done over
the last years.
Our goal is to formulate an optimization problem that expresses the need
of finding free parking slot in an efficient way, and then apply the ant colony
paradigm to solve that optimization problem.
3.1 Environment Classification
In this section, we want to classify the context of our problem [2]. Given the
problem of parking, the environment in which the problem can be imagined is
a huge area having taken or free parking slot. Our first assumption is to divide
the environment in regions, such that each region has a capability in terms of
available free parking slot.
The environment can be represented as a directed graph G = (V, E), where
each vertex v ∈ V is a region with |V | = n, and each edge (i, j) ∈ E is a
connection between regions, such (i, j) ∈ E if and only if from the region i the
region j is reachable directly. Moreover, a weight function w : V × V −→ N
associates to each edge a weight given by the capability of the destination node.
3.2 ACO Parameters

The main parameters in ant colony optimization paradigm are the following [25]:
• ηij is the attractiveness of the move from i to j, according to a priori infor-

mation;
• τij is the trail level of the move from i to j, according the a posteriori infor-
mation, given by previous iterations.
At any iteration, the ant k performs a move from i to j with a certain prob-
ability pkij that depends on attractiveness and trail level values. The probability
is computed as follows:
⎧ α β
⎪
⎨ τij + ηij
if (ij) ∈ tabuk
pkij = α + ηβ )
(τij ij
(1)
⎪
⎩
(ij) ∈ tabuk
0 otherwise
where:
• α and β are parameters between 0 and 1 that allow to specify how determinant
are the trail level and the attractiveness, respectively;
• tabuk is the list of not feasible moves for the ant k starting from the state i.
Once any ant completes its move, the trail level at iteration t is updated
according to the trail level at the previous iteration, as follows:
τij (t) = ρ ∗ τij (t − 1) + Δτij (2)
where: ρ is a parameter representing the evaporation rate that guarantees a way

to avoid too fast convergence of the algorithm to a suboptimal region, allowing
the exploration of new areas of the search space [20]; Δτij is the pheromone,
meaning the sum of the contributes given by all the other ants that have per-
formed the ij move already.
3.3 ACO Algorithm

Typically, an algorithm following ant colony optimization is based on the repe-
tition of three phases [20] until a termination condition is reached, as shown in
Fig. 1.
The initialization phase consists in assigning a starting value to trail level
and pheromone parameters, for any edge of the graph.
The construction phase is about the colony of ants, each of them comput-
ing its own solution according to a stochastic local decision policy, based on
pheromone and trail level.
Fig. 1. Phases of a typical algorithm solving a problem according to ant colony opti-
mization
The updating phase is the moment when trail level and pheromone are mod-
ified: in particular, the trail level can either increase or decrease according to if
ants deposit pheromone or not.
3.4 Parking Optimization Problem

To formulate the optimization problem that represents the situation where a
driver moving toward a destination wants to find a free parking slot, we need to
take into account:
• The position of the driver;
• The position of the destination.
Clearly, the ideal situation would be finding the parking slot which is the
closest one to the destination point. Our purpose is to solve the problem by
finding a path, from the position where the driver is to the one of the place he
wants to reach, that maximizes the number of free parking slot met. Figure 2a
shows an example of a graph made by five regions connected according to the
free parking slot available in the destination region, while Fig. 2b shows the path
that maximizes the number of available parking slot met, assuming that the
vehicle start from region A, and wants to reach region E.
In terms of the graph we have built in Subsect. 3.1, this corresponds to find a
path, from the region S where the driver starts to the region E where the driver
has to go to, with maximum weights over the edges.
Formally, let wij be the weight of the edge i → j in the graph G, and let
P be the resulting path of maximum weights, the optimization problem can be
formulated as follows: ⎧
n n
⎪
⎪
⎪
⎪ max wij ∗ xij
⎪
⎪
⎪
⎪ i=1 j=1
⎪
⎪ n
⎪
⎪
⎨ xij ≤ 1
(3)
⎪
⎪
i=1
⎪
⎪ n
⎪
⎪
⎪
⎪ xij ≤ 1
⎪
⎪
⎪
⎪ j=1
⎩
P [1] = S
where xij can assume values in [0, 1] according to the resulting path P :

1 if i → j ∈ P
xij =
0 otherwise
Fig. 2. Example of a graph of five regions and corresponding weighted connections (a),
and the chosen maximizing free parking slot path assuming A and E as starting and
ending point respectively (b).
The first inequality constraint guarantees that each node of the resulting
path has at most one outgoing arc (except for E that has not outgoing arcs).
Similarly, the second inequality constraint ensures that any node has at most
one incoming arc (zero in case of S). Notice that in the optimization problem
shown in (3), it is required that S appears in the resulting path (as first region
precisely), but this does not happen for E: indeed, if the destination region has
not free slot, any edge leading to it has a null weight and hence it is not chosen
for the optimal path.
4 Conclusions
Ant colony optimization is largely used to solve known problems, such as TSP
(Traveling Salesman Problem), SOP (Sequential Ordering Problem), GAP (Gen-
eralized Assignment Problem), and MKP (Multiple Knapsack Problem) [20]. In
this work, we turn the free parking slot detection problem into an optimization
problem, with the aim of finding a feasible solution that maximize the objective
function. We first build an opportune environment representation, then we for-
mulate the optimization problem to find a path from a starting position, leading
as close as possible to a destination, that maximizes the number of free parking
slot seen. As far as we know, this is the first work relating ACO optimization
with parking problem.
References
1. Aadil, F., Bajwa, K.B., Khan, S., Chaudary, N.M., Akram, A.: CACONET: ant
colony optimization (ACO) based clustering algorithm for vanet. PloS One 11(5),
e0154080 (2016)
2. Amato, F., Boselli, R., Cesarini, M., Mercorio, F., Mezzanzanica, M., Moscato, V.,
Persia, F., Picariello, A.: Challenge: processing web texts for classifying job offers,
pp. 460–463 (2015). Cited By 16
(2016). Cited By 18
4. Amato, F., Mazzocca, N., Moscato, F.: Model driven design and evaluation of
security level in orchestrated cloud services. J. Netw. Comput. Appl. 106, 78–89
(2018). Cited By 2
5. Amato, F., Moscato, F.: Model transformations of mapreduce design patterns for
automatic development and verification. J. Parallel Distrib. Comput. 110, 52–59
(2017). Cited By 3
6. Balzano, W., Murano, A., Vitale, F.: V2V-en-vehicle-2-vehicle elastic network. Pro-
cedia Comput. Sci. 98, 497–502 (2016)
7. Balzano, W., Murano, A., Vitale, F.: WiFACT–wireless fingerprinting automated
continuous training. In: Proceedings of WAINA. IEEE Computer Society (2016)
9. Balzano, W., Del Sorbo, M.R., Murano, A., Stranieri, S.: A logic-based clustering
approach for cooperative traffic control systems. In: 3PGCIC. Springer (2016)
10. Balzano, W., Vitale, F.: Dig-park: a smart parking availability searching method
using V2V/V2I and DGP-class problem. In: Proceedings of the WAINA. IEEE
Computer Society (2017)
11. Balzano, W., Barbieri, V., Riccardi, G.: Car2Car framework based on DDGP3
12. Balzano, W., Del Sorbo, M.R., Stranieri, S.: A logic framework for C2C network
management. In: 2016 30th International Conference on Advanced Information
Networking and Applications Workshops (WAINA), pp. 52–57. IEEE (2016)
pression of geodetic tracks. In: International Journal of Grid and Utility Computing
(2017)
15. Balzano, W., Stranieri, S.: LoDGP: a framework for support traffic information
systems based on logic paradigm. In: International Conference on P2P, Parallel,
Grid, Cloud and Internet Computing, pp. 700–708. Springer (2017)
16. Balzano, W., Vitale, F.: DGP application for support traffic information systems
in indoor and outdoor environments. In: International Conference on P2P, Parallel,
17. Balzano, W., Vitale, F.: PAM-SAD: ubiquitous car parking availability model
based on V2V and smartphone activity detection. In: International Conference
on Intelligent Interactive Multimedia Systems and Services, pp. 232–240. Springer
(2017)
18. Chauhan, R.K., Dahiya, A.: AODV extension using ant colony optimization for
scalable routing invanets. J. Emerg. Trends Comput. Inf. Sci. 3(2), 241–244 (2012)
19. Da Cunha, F.D., Boukerche, A., Villas, L., Viana, A.C., Loureiro, A.A.: Data
communication in VANETs: a survey, challenges and applications. Ph.D. thesis,
INRIA Saclay; INRIA (2014)
20. Dorigo, M., Birattari, M., Blum, C., Clerc, M., Stützle, T., Winfield, A.: Ant
Colony Optimization and Swarm Intelligence: Proceedings of the 6th Interna-
tional Conference, ANTS 2008, Brussels, Belgium, 22–24 September 2008, vol.
5217. Springer (2008)
21. Jindal, V., Dhankani, H., Garg, R., Bedi, P.: MACO: modified ACO for reducing
travel time in VANETs. In: Proceedings of the Third International Symposium on
Women in Computing and Informatics, pp. 97–102. ACM (2015)
22. Jung, H.G., Kim, D.S., Yoon, P.J., Kim, J.: Parking slot markings recognition for
automatic parking assist system. In: 2006 IEEE Intelligent Vehicles Symposium,
pp. 106–113. IEEE (2006)
23. Jung, H.G., Lee, Y.H., Kim, J.: Uniform user interface for semiautomatic parking
slot marking recognition. IEEE Trans. Veh. Technol. 59(2), 616–626 (2010)
24. Majumdar, S., Rajendra Prasad, P., Santosh Kumar, S., Sunil Kumar, K.N., et al.:
An efficient routing algorithm based on ant colony optimisation for VANETs. In:
IEEE International Conference on Recent Trends in Electronics, Information and
Communication Technology (RTEICT), pp. 436–440. IEEE (2016)
25. Maniezzo, V., Gambardella, L.M., De Luigi, F.: Ant colony optimization (2004).
Cited in New Optimization Techniques in Engineering. Studies in Fuzziness and
Soft Computing, vol. 141. Springer, Heidelberg, p. 62 (2010)
26. Suhr, J.K., Jung, H.G.: Sensor fusion-based vacant parking slot detection and
tracking. IEEE Trans. Intell. Transp. Syst. 15(1), 21–36 (2014)
27. Zeadally, S., Hunt, R., Chen, Y.-S., Irwin, A., Hassan, A.: Vehicular ad hoc net-
works (VANETs): status, results, and challenges. Telecommun. Syst. 50(4), 217–
241 (2012)
A Self-organization Technique in Wireless
Sensor Networks to Address Node
Crashes Problem and Guarantee Network
Connectivity

Abstract. Wireless Sensor Networks (WSN) are largely employed to

collect and elaborate data and information in a given environment. These
networks are made by power-constrained sensors able to receive and
transmit data wireless. Typically, this information is gathered by a sin-
gle sensor which has the responsibility of elaborating it, and inferring
something about the environment. One of the most desirable features
for a WSN is the fault tolerance. Because of the limited energy of the
sensors, node crashes may happen in the network, and this shouldn’t
affect the connectivity of the network itself. The fault tolerance prop-
erty is related to self-organizing capability that a WSN is supposed to
have, and that is often obtained through network clusterization. In this
work, we want to address the fault tolerance problem together with self-
organizing requirement, in order to provide a network satisfying both
robustness and autonomy needs. To this aim, we propose a clustering
algorithm that helps to preserve the network connectivity after a node
crash.
Keywords: Self-organizing · Clustering · Fault-tolerance
1 Introduction
Wireless Sensor Networks (WSN) have lately attracted the attention of
researchers from different research fields, where modeling aimed to security
improvement are hot topics [2–4], due to their potential and the numerous appli-
cation areas, such as traffic control [14,16], road safety [7,11], object classification
[1] and tracking [5,13], environment monitoring [8,9,23], but also surveillance
[19] and parking [12].
This kind of networks have two main limitations: (i ) energy-constrained sen-
sors, and (ii ) limited communication range. In order to optimize the use of the
restricted bandwidth, a smart message transmission system is realized: a sen-
sor in a WSN is allowed to communicate only with its one-hop neighbor, while
long-range transmission are obtained by multi-hop communication, letting all
https://doi.org/10.1007/978-3-030-15035-8_82
the sensors cooperate to execute distributed tasks. Shorabi, in [26], gives a hier-
archical interpretation of a wireless sensor network, allowing to process data
and information at different abstraction levels: each event occurring in the net-
work environment can be processed at (i ) node level, meaning from the sensor
detecting the event point of view, (ii ) local neighborhood level, meaning from
the one-hop neighbors of the sensor detecting the event point of view, and (iii )
global level, which involves the whole network.
Sensors in WSN typically deal with highly dynamic environment: changes
in the weather conditions, occurrence of exceptional events, error detection are
some of the possible causes that determine rapid and frequent changes in the
environment conditions. In these cases, a WSN is supposed to be autonomous
enough to self-manage, without any human intervention [24]. Essentially, a WSN
is required to be fault-tolerant.
The ability of a WSN of solving problems after exceptional events by itself is
known as self-organizing property, meaning the autonomous organization of con-
nectivity, and data addressing. Self-organization is a challenge in wireless sensor
networks, because of the mentioned limitations like restricted energy and band-
width. The power constraint, especially, poses the problem of fault-tolerance:
when a sensor shuts down, because of energy expiration, the connectivity of the
whole network must not be affected by the failure of a single node.
In order to address the constrained energy of sensors, clustering techniques
are often employed to reduce power waste. In a typical scenario, data retrieved by
each sensor in a WSN are transmitted to central processing units, that gather all
the information and communicate them outside the network. As they point out
in [15], the cost for transmitting information is often higher than the one needed
for the computation. This fact suggests using clustering approaches. Indeed,
intuitively dividing sensors in groups reduces the distances that a single node
has to cover, because each sensor has to communicate its own information to
a representative node of its group, known as clusterhead, which is closer, in
terms of distances, than the central processing unit. The clusterheads of each
group receive data from all the other related sensors, and are responsible for
the transmission of such data to the central units. This mechanism reduces the
energy waste significantly for the whole network.
As we show in the next section, researchers have proposed different clustering
algorithms to optimize transmission in wireless sensor networks, and to make
the most of the available resources. Some of them also focused on sensors power
issue, providing energy-based clustering algorithms, but very few addressed the
problem of WSN fault-tolerance. The main contribution of this paper is a fault-
tolerant clustering algorithm, aiming to preserve the network connectivity even
after node crashes, and to improve the self-organizing ability.
Outline
The rest of the work is organized as follows: in Sect. 2, the state of the art is
analyzed and the existing clustering techniques are shown; in Sect. 3, a clustering
algorithm which is resistant to faults and node crashes is proposed, with relative
detection and recovery phases; finally, in Sect. 4, the conclusions of this work.
A Self-organization Technique in Wireless Sensor Networks 843
2 Related Work
One of the hot topics in Wireless Sensor Networks field is the self-organization
problem: as the authors point out in [24], WSNs should be able to manage
themselves and to work without any human intervention. Flako Dressler in [17]
explains that self-organization is a paradigm to handle distributed systems: this
property is required when the system need to face a change in the environment
circumstances, or to adapt to new scenarios, features that allow to improve
scalability and adaptivity in any control system.
In [26], the authors provide a survey explaining the existing self-organization
techniques, with related progress and open issues. They highlight that the WSN
ability to self-organize has been emphasized to address the problem of managing
large-scale networks, and has been proven to bring significant benefits in coop-
erative systems performances. On the other hand, self-organization requires a
new control theory to realize such an autonomous system.
For this reason, many researchers propose clustering as control strategy to
achieve such a self-managing ability. In our previous work, we focus on clus-
terization too [6,10]. In [22], they propose a load-balanced clustering algorithm
based on distance and density distribution (DSBCA). The aim of this algorithm
is to guarantee the generation of clusters that have balanced energy and to avoid
the situation with overcrowding clusters. Essentially, they propose to compute
the clustering radius based on density and distance such that when two clusters
have same density, the much further one from the base station takes a larger
radius, and when two clusters have same distance from the base station, the one
having higher density takes a smaller radius.
In [25], they exploit the fact that in a wireless sensor network communication
is the most consuming operation in terms of energy and they provide a set of
algorithms for keeping the network connected: these algorithms are based on
protocols for self-organization.
Authors in [21] show three different approaches to achieve self-organizing
property, by exploiting clustering algorithms to create groups of fixed size B : (i )
in the Expanding Rapid Alorithm, an initiator node includes in its cluster the
sensors in a neighborhood that increases at each iteration, until the cluster-size
is reached (some nodes of the last level may be cut off in case B is exceeded);
(ii ) in the Rapid Algorithm, the initiator, instead, is assigned a budget (given by
the cluster-size) which is first decremented and then equally distributed among
its one-hop-neighbors, that will repeat the procedure until the budget is expired
or there is no possibility to let the cluster grow: the main drawback is that this
algorithm may produce groups smaller than the resource would allow; (iii ) for
this reason, they propose a third algorithm, the Persistent Algorithm, which is
an extension of the previous one since it allows the nodes receiving the budget
to give back to the father the unused part of the budget, so that it can be
distributed again: this guarantees that the size of the groups is much more close
to the maximum cluster-size.
Moreover, in [15], they propose a randomized, distributed clustering algorithm
to manage sensors in WSNs and then they extend it to generate a hierarchy
of clusterheads. Each sensor has a probability of being clusterhead. They distin-

guish volunteer clusterheads from forced clusterheads: the former are sensors that
advertise themselves as clusterheads to all the sensors that are within a fixed dis-
tance; the latter are sensors that are neither clusterheads nor have joined existing
clusters within a fixed time.
Some researchers relates clustering algorithms to fault-tolerance problem,
trying to produce clusters in such a way that the recovery after an error is easier.
In [20], they point out that self-configuration and self-organization mechanisms
are needed to satisfy the requirement of operating in dynamic environment.
Together with this constraint, they also want to guarantee that faulty behavior
of a single sensor does not have any impact on the overall system behavior. For
this reason, they propose an algorithm based on the detection of event regions,
meaning network portions where interesting events occur, disambiguating faults
by events.
Furthermore, in [18], the authors provide a model based on two kind of nodes:
sensors and gateways. The difference is that the latter are less-energy-constrained
sensors. Their mechanism has a first phase of detection and a second phase of
recovery. The fault detection is obtained by periodic updates received by the
gateways from all the related sensors. When an update misses, a fault has been
occurred. The recovery starts as soon as this fault is detected and consists in
allocating the sensors to new clusters. During the clustering, each gateway has
a Bset of nodes included in its communication range, but not included in its
cluster, in such a way that when a sensor needs to be recovered, all the gateways
check their Bset and the sensor is eventually assigned to a new cluster.
3 Clustering Node Crashes-Resistant Network

Our approach is based on a representation of a Wireless Sensor Network as a
graph, as follows:
• The set V of vertices is the set of sensors;
• The set E of edges is built such that two vertices i and j are connected by
an edge if and only if they are close enough to communicate to each other.
The main goal of this approach is to guarantee network connectivity even after a
node fault, i.e. we want to avoid that sensors may be isolated from the network
if another sensor crashes. An example is provided in Fig. 1: let’s focus on the
blue cluster and let’s suppose that the node 40 is the clusterhead, and the node
47 can communicate with it only by multi-hop transmission through the node
48. Now, let’s suppose that the sensor 48 shuts down due to low battery: this
is an example in which a sensor of the cluster (the 47) remains isolated from it
because the node, that acted as bridge between it and the clusterhead, crashed.
This is one of the situation we want to avoid with our approach.
One of the main causes of sensor fault is due to low battery, since sensors
in a WSN are energy-constrained devices. To avoid energy waste in a WSN, the
first step suggested by the literature is the clusterization of the network. Indeed,
Fig. 1. Example of network made of 30 sensors arranged in 3 clusters
cluster creation allows a smart information distribution and data collection, since
there is a single node, elected as cluster-head, which is responsible for gathering
the information about all the other sensors in the same cluster, and transmitting
it outside the network. Clearly, dividing a network of hundreds of sensors in
groups of tens reduces not only the network congestion but also the effort made
by each node to collect all the data sent in the network.
In order to guarantee network connectivity, we propose a clustering algo-
rithm, based on backup-clusters, that helps the network be resistant with respect
to errors. A cluster is a connected component, meaning that there exists a path
between any pair of nodes, regardless of the cluster-head choice: this fact guar-
antees that, even after the cluster construction, the role of clusterhead might
change without impact the cluster connectivity.
3.1 Cluster Creation
Anytime a node appears in the network, it waits a random time t during which
it can be asked to join an existing cluster. If this time expires without any
invitation, the node becomes a temporary cluster-head, and it invites its adjacent
nodes to join the cluster. Given the considerations made previously about cluster
connectivity, during the cluster composition the cluster-head role can be given
to another node if it has left more energy. This choice aims to facilitate the
recovery after a crash: clearly, if an error occurs in the cluster-head the recovery
is more difficult than it would be if any other node turns off, as we will explain
later.
Let’s suppose to have a fixed maximum cluster-size S : this size is used during
the cluster creation as criterion to stop the cluster expansion, but it will be
relaxed with a tolerance ε during the recovery.
During the clustering procedure, each node is assigned:
• a cluster id, which identifies its cluster, and thus the corresponding cluster-
head;
• a backup-clusters list, given by the cluster ids of adjacent nodes not belonging
to the same cluster of the current one. It would be appropriate keeping this list
ordered according to the backup-clusters emptiness: this requirement allows
to avoid overcrowding during the recovery, that would reduce significantly
the benefits of clusterization. In this way, the first item of the list contains
the most available backup-cluster id.
At any time a node is aware of its current cluster and, if a fault determining
its isolation from the rest of the cluster occurs, it knows exactly where it could
be reassigned preserving the network connectivity.
The Fig. 2 summarizes the steps of the approach proposed in this work:
• the cluster expansion is realized starting by the temporary clusterhead that

invites its one-hop neighbors to join the cluster. Its neighbors forwards the
invitation to their one-hop neighbors, and so on until the desired size is
reached, or no further expansion is possible.
• after the cluster construction, the temporary clusterhead receives information
about all the sensors belonging to its cluster, including the remaining battery.
Given this information, a new strategic clusterhead is elected, as the sensor
in the cluster with the highest energy left.
3.2 Backup Clusters
The flow-chart showed in Fig. 2 has to be enriched with an additional phase,

which is the backup-clusters list creation. Before defining how this list is made,
let’s first fix some notation.
Fig. 2. Flow-chart showing the steps from the time when a new node joins the network
to the cluster creation
Notation
• d(ni , nj ) gives the distances between the sensors ni and nj ;

• R is the maximum communication range, meaning the maximum distance
between two adjacent nodes;
• cluster id(n) gives the identifier the represents the cluster containing n;
• f ree space(id) gives the available space in the cluster identified by id,
obtained as the difference between the maximum cluster size S and the actual
cluster size.
Given the node N , its backup-clusters list is computed as the list of sensors
in its communication range, but not in its cluster:
backup-clusters(N ) = {n1 , . . . , nk }
such that ∀i ∈ {1, . . . , k}:

1. d(ni , N ) ≤ R
2. cluster id(N ) = cluster id(ni )
The list backup-clusters(N ) is kept ordered by f ree space(cluster id(ni )).
3.3 Fault Detection and Recovery
In this section, we show the network behavior after an error occurrence, when
our clusterization technique has been used. In order to solve a node crash, we
need to distinguish two phases: the detection and the recovery.
3.3.1 Detection
In this phase, the aim is to understand that a node crashed. A sensor in a
wireless sensor network may shut down because of low battery, or any kind of
fault. Whatever the cause is, we just want to find a way to establish that a
crash occurred. Differently from other proposals, we don’t focus on the crashed
node, but on the one that is affected by the fault. Indeed, we are not interested
in letting the clusterhead know which sensor crashed, but we want to let the
isolated node be aware of its isolation, so that it can be fixed.
In each cluster there is a periodic information exchange: in particular, we
assume a communication scheme such that the clusterhead receives updates by all
the sensors in its cluster at arbitrary time intervals, and it sends back a “received”
answer to all the nodes that have sent updates (both in one or multi hops).
When a fault occurs, if a sensor remains isolated from the rest of the cluster,
it does not receive the answer after sending its updates, and it automatically
detects the error.
3.3.2 Recovery
The recovery strategy, following a node crash in a network clusterized with the
algorithm explained above, is made simple by the backup-clusters list. The node
I, that finds itself in isolation from the rest of the cluster, can check its backup-
clusters list and join the cluster identified by the first item of the list, by broad-
casting a message containing its id and the id of the cluster to join. The neigh-
bor(s) receiving this message checks its cluster id and if it matches, the message
is propagated, via multi-hop if needed, to the new clusterhead that adds the new
node to its cluster.
Notice that if the crashed node is a clusterhead, a total reassignment occurs,
meaning that all the sensors in the clusters find themselves in isolation, and they
all apply this recovery technique, reassigning themselves to new clusters. In order
to avoid frequent total reassignments, we introduced in the previous section the
requirement of electing the clusterhead as the sensor with highest energy.
Algorithm 1. Reassignment algorithm

input : isolated node I
1 while REJECTED msg is received do
2 cluster id ← next(backup-clusters(I));
3 send(receiver=cluster id, msg=JOIN) ;
4 end
Algorithm 2. Reception of join request algorithm

input : applicant identifiers id, tolerance ε, max cluster size S
1 if this cluster cardinality less than S + ε then
2 send(receiver=id, msg=ACCEPTED);
3 else
4 send(receiver=id, msg=REJECTED);
5 end
A possible drawback could be the excessive growth of clusters already existing

after crashes: indeed, the fixed maximum cluster size S can be easily exceeded.
For this reason, we fix a parameter ε that allows to relax the cluster size con-
straint, such that the cluster size can be up to S + ε. In this way, when a sensor
tries to join a new cluster after a crash, its request can also be rejected if the
cluster size exceeded (this is needed in order not to lose the benefits of cluster-
ization with overflowing clusters). The rejected node can try to join the cluster
identified by the next item of its backup-clusters list, until it is reassigned.
The pseudo-code in 4 shows how an isolated node tries to reassign itself to
another cluster in its backup list, while the pseudo-code in 5 shows the cluster-
head point receiving the request point of view: it accept the new node only if
there is enough space in its cluster.
4 Conclusions
In this paper, we address the problem of improving the self-organizing capability
of a wireless sensor network. In particular, we guarantee the network connectivity
even after exceptional errors or faults of a single sensor. The self-organization
property is obtained by exploiting clustering techniques, that allow the network
to minimize the flow of messages, and the communication between nodes, which
constitute the most expensive operation in terms of energy. We want to avoid the
situation where, because of a sensor crash, another sensor cannot communicate
with its cluster anymore. For this reason, provide a clusterization algorithm
letting each sensor keep track of a backup list of clusters, to which it can be
reassigned in case it remains isolated in its cluster.
References
pp. 460–463 (2015). Cited By 16
of multimedia data for E-government applications. J. Vis. Lang. Comput. 32, 35–41
(2016). Cited By 18
(2018). Cited By 2
4. Amato, F., Moscato, F.: Model transformations of mapreduce design patterns for
(2017). Cited By 3
5. Arora, A., Dutta, P., Bapat, S., Kulathumani, V., et al.: A line in the sand: a
wireless sensor network for target detection, classification, and tracking. Comput.
Netw. 46(5), 605–634 (2004)
7. Balzano, W., Murano, A., Vitale, F.: V2V-EN-vehicle-2-vehicle elastic network.
Procedia Comput. Sci. 98, 497–502 (2016)
management. In: Proceedings of WAINA. IEEE Computer Society (2016)
using V2V/V2I and dgp-class problem. In: Proceedings of WAINA. IEEE Com-
puter Society (2017)
pression of geodetic tracks. Int. J. Grid Util. Comput. (2017)
14. Balzano, W., Stranieri, S.: Cooperative localization logic schema in vehicular ad
hoc networks. In: International Conference on NBis, pp. 960–969. Springer (2018)
15. Bandyopadhyay, S., Coyle, E.J.: An energy efficient hierarchical clustering algo-
rithm for wireless sensor networks. In: Proceedings of IEEE INFOCOM 2003. IEEE
(2003)
16. Chen, W., Chen, L., Chen, Z.-l., Tu, S.-l.: A realtime dynamic traffic control system
based on wireless sensor network. In: 34th International Conference on Parallel
Processing Workshops. IEEE Computer Society (2005)
17. Dressler, F.: Self-organization in Sensor and Actor Networks. Wiley, Chichester
(2008)
18. Gupta, G., Younis, M.F.: Fault-tolerant clustering of wireless sensor networks. In:
IEEE WCNC 2003, New Orleans, 16–20 March 2003. IEEE (2003)
19. He, T., Krishnamurthy, S., Stankovic, J.A., et al.: Energy-efficient surveillance
system using wireless sensor networks. In: Proceedings of the 2nd International
Conference on Mobile Systems, Applications, and Services, pp. 270–283. ACM
(2004)
20. Krishnamachari, B., Iyengar, S.: Distributed Bayesian algorithms for fault-tolerant
event region detection in wireless sensor networks. IEEE Trans. Comput. 53(3),
241–250 (2004)
21. Krishnan, R., Starobinski, D.: Efficient clustering algorithms for self-organizing
wireless sensor networks. Ad Hoc Netw. 4(1), 36–59 (2006)
22. Liao, Y., Qi, H., Li, W.: Load-balanced clustering algorithm with distributed self-
organization for wireless sensor networks. IEEE Sens. J. 13(5), 1498–1506 (2013)
23. Mainwaring, A., Culler, D., Polastre, J., Szewczyk, R., Anderson, J.: Wireless sen-
sor networks for habitat monitoring. In: Proceedings of the 1st ACM International
Workshop on Wireless Sensor Networks and Applications, pp. 88–97. ACM (2002)
24. Sharma, S., Bansal, R.K., Bansal, S.: Issues and challenges in wireless sensor net-
works. In: International Conference on Machine Intelligence and Research Advance-
ment (ICMIRA), pp. 58–62. IEEE (2013)
25. Sohrabi, K., Gao, J., Ailawadhi, V., Pottie, G.J.: Protocols for self-organization of
a wireless sensor network. IEEE Pers. Commun. 7(5), 16–27 (2000)
26. Zhang, Z., Long, K., Wang, J.: Self-organization paradigms and optimization
approaches for cognitive radio technologies: a survey. IEEE Wirel. Commun. 20(2),
36–42 (2013)
Data Dissemination in Vehicular Ad Hoc
Network: A Model to Improve Network
Congestion

Abstract. Network congestion is a serious problem affecting high den-

sity network. In vehicular context, the traffic conditions and the huge
amount of messages sent by vehicles make this problem concrete. The
more the network congestion increases, the more the VANET (Vehic-
ular ad Hoc Network) safety is put at risk. Typically, this problem is
faced by acting a congestion detection followed by a congestion control
strategy that schedules the messages transmission using some metrics to
decide which messages have to be assigned the highest priority. By ana-
lyzing vehicles behavior in VANETs, we observed that the very propaga-
tion mechanism behind inter-vehicular communication allows redundant
transmissions, that might favor network congestion. For this reason, in
this work, we want to propose a starting preliminary phase allowing the
redundancy detection and, hence, the congestion prevention. Following
this approach, the congestion control strategies need to be applied less
frequently than before. By means of Petri Net modeling language, we
explain the behavior of a vehicle adopting the proposed strategy.
Keywords: Congestion · Prevention · VANET · Petri Net
1 Introduction
Imagine a driver, hurrying for an important job meeting, being slow down
because of a road accident along the road he is cruising. What if the driver
had been previously warned about the accident? For sure, he could have chosen
a different path getting to his meeting on time, as well as the road safety would
have been improved. Inter-vehicular communication is thought to solve these
situations. VANETs (Vehicular ad Hoc Networks) are made of vehicles form-
ing a decentralized network, and communicating via On-board units (OBUs).
This kind of networks are based on short-range communication between vehi-
cles: neighbors can communicate directly, while the information propagation is
realized through multi-hop communication.
Given the numerous application fields, such as parking [10,11,16], network
management [7] and classification [2], cluster-oriented communication [6,9], com-
munication optimization [8,13–15], and since the number of vehicles equipped
https://doi.org/10.1007/978-3-030-15035-8_83
with wireless communication devices is growing, inter-vehicular communication

is becoming a promising research field. Indeed, VANETs not only can improve
road safety, but also can be employed to compute distributed tasks. In such a
network, communication is based on broadcast transmission, that can be based
on two different mechanisms, as they point out in [28]: (i ) flooding and (ii )
dissemination. In the first case each vehicle broadcasts information about itself
periodically; when a vehicle receives a message, it is forwarded immediately. In
the second case, each vehicle broadcasts information about itself and other known
vehicles; when a vehicle receives a message, it updates its own information, that
are forwarded at the next broadcast period. They also assume a direction in
vehicle propagation, and messages propagation, by distinguish three dissemina-
tion models, (i ) same direction, (ii ) opposite direction, and (iii ) bidirectional,
according to which vehicles are responsible for messages propagation.
As pointed out in [31], a key concern in this field is the system scalability.
Unfortunately, scalability may be affected by a common phenomenon in high
density network, which is congestion. VANETs dynamic environment, changing
topology, and distributed nature make congestion control more challenging, as
they also highlight in [22]. In [30] they say that to avoid congestion, the channel
load can be reduced and controlled by a dynamic adaptation of transmission
power, packed transmission frequency, and packet duration. Due to the limited
bandwidth in multi-hop inter-vehicular communication, many congestion control
strategies proposed by researchers aim to favor safety messages transmission, by
assigning them priorities.
In this work, we show how redundant transmission impact network conges-
tion in VANETs, and we propose a redundancy avoidance phase preceding the
congestion control one (if needed), in order to reduce the overall network conges-
tion. We use a modeling approach [5], relying on Petri Net modeling language,
whose semantic [3] is used to explain the behavior of such a system.
Outline
The rest of the paper is organized as follows: in Sect. 2, a comparison between
proposals made by other researchers about congestion control is performed; in
Sect. 3, our proposal is provided, by analyzing the impact of redundancy of
network congestion, and by proposing a solution to handle it; finally, in Sect. 4,
we provide the conclusions of this work.
2 Related Work
VANET constitutes a recently increasing research area aimed to exploit vehicular

communication for road and drivers safety purposes [4,18]. Among the numerous
open issues reported in [34], they indicate the congestion control as a challeng-
ing topic in this field, despite being largely studied. This is a technique aimed
to handle network overload in case of high density configuration. In literature,
we can find different definitions of congestion, such as in [23], it is indicated
as a user-visible degradation of network performance, or in [32], they speak
Data Dissemination in Vehicular Ad Hoc Network 853
about congestion as a phenomenon occurring when resource demands exceed

the capacity, or again in [22] they define congestion as the situation when the
rate of injected packets exceeds its processing capacity over a continuous period
of time leading to and increasing number of packets loss. In general, conges-
tion causes the inability of the network to process all the requests, and hence it
needs to be handled, especially in vehicular context, where high density is quite
common.
Different control strategy have been proposed, such as in [22], they address
the network congestion issue in VANETs by proposing a solution made of three
phases: (i ) priority assignment to messages, (ii ) congestion detection, and (iii )
adjustments. The first phase is needed when a vehicle receives more than one
message: a priority is assigned to each message according to its degree of impor-
tance and danger. The second phase allows vehicles to detect congestion, based
on some metrics characterizing the network. When a congestion occurs, the third
phase involves the transmit power and beacon transmission rate adjustment to
facilitate emergency messages spread.
In [20], in order to schedule messages transmission, they also rely on priorities
evaluated as a function of the utility of the message, the sender application,
and the neighborhood context. In [33], collision warning takes the priority: they
focus on collision avoidance, by proposing a communication protocol aimed to
guarantee the delivery of collision warning messages.
In [31], they present a design methodology for congestion control in Vehic-
ular Safety Communication, while [19] provides two approaches (reactive and
adaptive) based on transmit message rate control. For an overview of existing
congestion control protocols, consult [25].
Congestion avoidance has been investigated also in situations where different
agents need to reach a certain area without collide to each other [29], or even
for prediction purposes [26].
Some researchers focus on congestion prevention, rather than congestion con-
trol, such as in [30], where they rely on cooperation to prevent channel conges-
tion by exploiting traffic information to reduce interferences and channel load.
Cooperation is also used in our recent work in [17] with localization purposes.
In this work, we propose a prevention technique, preceding the control one
if needed, aimed to reduce transmission redundancy. Indeed, we observed that
in a vehicular network, due to high density, each vehicle can receive the same
message from different directions. This is a situation we want to avoid, since
the evaluation of information, that is already known, is a waste of time and
resources. To this aim, we exploit a largely employed modeling language for the
representation of distributed systems behavior: Petri Net. This directed bipartite
graph is commonly used in VANET field. Indeed, in [24], they use continuous
PN as a model to design traffic network behavior and analyze traffic flow; in
[21], similarly to the previous case, they employ hybrid PN with the purpose
of solving the problem of coordinating several traffic lights aimed to favor some
classes of vehicles, such as the emergency ones; one of our previous works, [12],
uses PN to describe a set of road exceptional events and how they can be handled
through Car2Car communication. In this work, Petri Net are used to represent
the behavior of a VANET aimed to reduce redundant computations, so to prevent
network congestion.
3 Redundancy Impact Elimination for Congestion

Prevention
Before analyzing the details of the approach proposed in this work, we fix some
basic notations about how to interpret a vehicular network.
Definition 1. A VANET can be seen as a directed graph G = <V, E>, where
the set of vertices V is the set of vehicles, while the set of edges E is such that
the pair (u, v) ∈ E if and only if v is in the communication range of u, in terms
of RSSI (received signal strength indicator).
As we already mentioned, network congestion has been defined in different ways
by the researchers. It is essentially a network condition in which it is no longer
possible to accomplish all the incoming requests. In a vehicular context, network
congestion may happen in situations of high traffic density: for instance, when
a vehicle receives different messages coming from different neighbors, it has to
schedule their transmission according to some metrics. This is what researchers
do by means of congestion control mechanisms, often, as we have mentioned,
by associating priorities to messages. Without loss of generality, in this work we
assume that vehicles move from west to east and, hence, the message propa-
gation flows from east to west. We want to study the network congestion from
another perspective with respect to other researchers. Indeed, we observed that
redundancy in transmission may impact the congestion for the receiving vehicle
and, as a consequence, for the whole network. As Fig. 1 shows, if some event
occurs along the flow direction, the A vehicle would detect it before the others
and would transmit the message to its neighbors in the propagation direction: B
and C. These two vehicles would in turn propagate the information to D, that
would receive the same message twice.
Fig. 1. Example of redundancy transmission in a VANET

Congestion control mechanisms keep being undeniable to guarantee network

scalability and reliability. Our aim is not to substitute congestion control with
congestion prevention, but in our opinion this is a preliminary phase needed
to reduce the computation in congestion control phases, through redundancy
detection (as shown in Fig. 2).
Fig. 2. Integration between congestion prevention and control
3.1 The Analysis
We made some simulations using a known language for statistical computations,

R [1], to study the impact of redundancy on network congestion. Given the
number of nodes and a distance indicating the range within which vehicles can
see each other, we generated random networks. Then, by choosing a random
node as source of a new message dissemination, we simulated the propagation in
the opposite with respect to the direction of travel to estimate how many times
a vehicle could receive the same message, if the redundancy is not considered.
What came up is summarized in Fig. 3, where:
• the x axis represents the number of nodes of the network;

• the y axis is the number of the observed repetitions in our simulations;
• the solid line represents the observations without redundancy elimination;
• the dashed line is the ideal situation without redundancy: each node receive
a message once.
Clearly, the higher the number of nodes is the higher the impact of redundancy
is, since the network gets more congested. Moreover, in a simulation with only 40
nodes in the network, the number of repetitions of the same message overcomes
300, in case of very connected networks. If the network becomes made by hun-
dreds of nodes, the redundancy can seriously impact the VANET safety. This is
a consequence of the fact that, without any operation aimed to redundancy con-
trol, a node receives the same message as many times as the number of different
paths, following the propagation direction, from the source of the message to the
node itself. These observations let evident the need of redundancy elimination
mechanism beyond that congestion control one.
Fig. 3. Impact of redundancy on network congestion
3.2 The Model

Before defining the model representing the redundancy reduction phase, let us
make some assumption on the messages that vehicles exchange. In this work, we
assume that each vehicle can be uniquely identified by an id, and that a message
is characterized by the following information:
• a timestamp identifying the moment in which the message has been pro-
duced for the first time;
• a source indicating the id of the vehicle that produced the message for the
first time;
• a brief explanation that is the very content of the message representing
some situation of the network.
Thanks to the above information, we can conclude that two messages with same
timestamp and source are the same, hence one of them can be safely discarded.
According to the previous example, the D vehicle in Fig. 1 can exploit this knowl-
edge in order to discard one of the two messages (that are actually the same)
coming from its neighbors. In particular, the discarded message is the one coming
later.
In order to explain the behavior of such a system, we employ a largely used
modeling language for description of distributed systems, which is Petri Net,
being defined as follows [27].
Definition 2. A Petri Net is a 5-tuple P N = (P, T, F, W, M0 ), where P is a
finite set of places, T is a finite set of transition, F is the set of arcs linking
places to transitions and vice versa, W is the weighting function, and M0 is the
initial marking.
Definition 3. A marking is a function associating to each place a natural num-
ber representing the number of tokens in it. Typically, it is indicated as a vector
having in the i-th position the number of tokens of the i-th place.
The Petri Net in Fig. 4, with initial marking M0 = [1 0 0 0 0 0], represents the
behavior of a single vehicle receiving a new message from the network. A vehicle
Fig. 4. Petri Net representing the redundancy reduction phase
simply discards known messages, allowing a clever use of the available resources
that are employed to process and propagate only new messages. This simple
preliminary step could seriously improve the network scalability, since, as we have
seen from the graphics in Fig. 3, redundancy in a high density network can lead to
a very high number of repeated messages. Clearly, congestion control strategies
are still needed: indeed, it may happen that more events occur in areas of the
network which are far from each other, but their propagation converges to the
same node that receives more messages than allowed according to its resources. In
this case, even if redundancy elimination has been previously applied, congestion
control mechanism is required.
4 Conclusions
Network congestion is a big deal in VANET research field. Many congestion
control proposals can be found in literature, but the redundancy contribution
to this phenomenon has not had the same attention. For this reason, in this
work we provide an analysis of the impact of communication redundancy on the
network congestion problem, that highlights the need to manage this problem,
even before the congestion occurs. To this aim, we propose a model, exploiting
the Petri Net modeling language, that explains the behavior of a generic vehicle
of the network when a message is received, and the way the propagation is
handled.
References
1. R-Project: What is R? https://www.r-project.org/about.html
pp. 460–463 (2015). Cited By 16
of multimedia data for E-government applications. J. Vis. Lang. Comput. 32, 35–41
(2016). Cited By 18
(2018). Cited By 2
5. Amato, F., Moscato, F.: Model transformations of MapReduce design patterns for
(2017). Cited By 3
7. Balzano, W., Murano, A., Vitale, F.: V2V-EN-vehicle-2-vehicle elastic network.
8. Balzano, W., Murano, A., Vitale, F.: WiFACT-wireless fingerprinting automated
using V2V/V2I and DGP-class problem. In: Proceedings of WAINA. IEEE Com-
puter Society (2017)
11. Balzano, W., Barbieri, V., Riccardi, G.: Smart priority park framework based on
DDGP3. In: 32nd International Conference on Advanced Information Networking
and Applications Workshops (WAINA), pp. 674–680. IEEE (2018)
12. Balzano, W., Rosaria, M., Sorbo, D., Stranieri, S.: A logic framework for C2C
network management. In: 30th International Conference on Advanced Information
Networking and Applications Workshops (WAINA), pp. 52–57. IEEE (2016)
13. Balzano, W., Formisano, M., Gaudino, L.: WiFiNS: a smart method to improve
positioning systems combining WiFi and INS techniques. In: International Con-
ference on Intelligent Interactive Multimedia Systems and Services, pp. 220–231.
Springer (2017)
pression of geodetic tracks. In: The International Conference on Data Compression,
Communication, Processing and Security, CCPS 2016 (2016)
16. Balzano, W., Stranieri, S.: LoDGP: a framework for support traffic information
systems based on logic paradigm. In: International Conference on P2P, Parallel,
hoc networks. In: International Conference on Network-Based Information Sys-
tems, pp. 960–969. Springer (2018)
18. Balzano, W., Stranieri, S.: A logic user-based algorithm to improve node distribu-
tion in wireless sensor network. In: VLSS 2018. Springer (2018)
19. Bansal, G., Cheng, B., Rostami, A., Sjoberg, K., Kenney, J.B., Gruteser, M.:
Comparing LIMERIC and DCC approaches for VANET channel congestion con-
trol. In: IEEE 6th International Symposium on Wireless Vehicular Communica-
tions (WiVeC), pp. 1–7. IEEE (2014)
20. Bouassida, M.S., Shawky, M.: On the congestion control within VANET. In: Wire-
less Days, WD 2008. 1st IFIP, pp. 1–5. IEEE (2008)
21. Di Febbraro, A., Giglio, D., Sacco, N.: Urban traffic control structure based on
hybrid petri nets. IEEE Trans. Intell. Transp. Syst. 5(4), 224–237 (2004)
22. Djahel, S., Ghamri-Doudane, Y.: A robust congestion control scheme for fast and
reliable dissemination of safety messages in VANETs. In: Wireless Communications
and Networking Conference (WCNC), pp. 2264–2269. IEEE (2012)
23. Fowler, H.J., Leland, W.E., Bellcore, B.: Local area network traffic characteristics,
with implications for broadband network congestion management. IEEE J. Sel.
Areas Commun. 9(7), 1139–1149 (1991)
24. Julvez, J.J., Boel, R.K.: A continuous petri net approach for model predictive
control of traffic systems. IEEE Trans. Syst. Man Cybern. Part A Syst. Hum.
40(4), 686–697 (2010)
25. Mughal, B.M., Wagan, A.A., Hasbullah, H.: Efficient congestion control in VANET
for safety messaging. In: International Symposium in Information Technology
(ITSim), vol. 2, pp. 654–659. IEEE (2010)
26. Murano, A., Perelli, G., Rubin, S.: Multi-agent path planning in known dynamic
environments. In: PRIMA 2015. LNCS, vol. 9387, pp. 218–231. Springer (2015)
27. Murata, T.: Petri nets: properties, analysis and applications. Proc. IEEE 77(4),
541–580 (1989)
28. Nadeem, T., Shankar, P., Iftode, L.: A comparative study of data dissemination
models for VANETs. In: Third Annual International Conference on Mobile and
Ubiquitous Systems: Networking & Services, pp. 1–10. IEEE (2006)
29. Rubin, S., Zuleger, F., Murano, A., Aminof, B.: Verification of asynchronous
mobile-robots in partially-known environments. In: PRIMA 2015. LNCS, vol. 9387,
pp. 185–200. Springer (2015)
30. Sepulcre, M., Gozalvez, J., Härri, J., Hartenstein, H.: Contextual communications
congestion control for cooperative vehicular networks. IEEE Trans. Wirel. Com-
mun. 10(2), 385–389 (2011)
31. Tielert, T., Jiang, D., Chen, Q., Delgrossi, L., Hartenstein, H.: Design methodol-
ogy and evaluation of rate adaptation based congestion control for vehicle safety
communications. In: Vehicular Networking Conference (VNC), pp. 116–123. IEEE
(2011)
32. Welzl, M.: Network Congestion Control: Managing Internet Traffic. Wiley, Chich-
ester (2005)
33. Yang, X., Liu, L., Vaidya, N.H., Zhao, F.: A vehicle-to-vehicle communica-
tion protocol for cooperative collision warning. In: The First Annual Interna-
tional Conference on Mobile and Ubiquitous Systems: Networking and Services,
MOBIQUITOUS 2004, pp. 114–123. IEEE (2004)
34. Yousefi, S., Mousavi, M.S., Fathy, M.: Vehicular ad hoc networks (VANETs): chal-
lenges and perspectives. In: Proceedings of the 2006 6th International Conference
on ITS Telecommunications, pp. 761–766. IEEE (2006)
ManDri: A New Proposal of Manus VR
Facility Integration in Everyday Car
Driving
Walter Balzano(B) , Maurizio Minieri, and Silvia Stranieri

walter.balzano@gmail.com, maurizio.minieri@gmail.com,
silviastranieri1047@gmail.com
Abstract. The purpose of this paper is to analyze all the possible uses
of the Manus VR in everyday driving, to make it easier and safer. With
these technological gloves, equipped with eleven sensors for each hand,
able to precisely pick up every movement of the hand and fingers, it
is possible to associate various gestures to specific actions. In particu-
lar, they allow to turn, accelerate, brake, activate the car turn signals,
the air conditioning, even alarm, everything just using simple gestures.
But ManDri is not only framed in today’s guide: indeed, a glove, being
equipped with two IMUs (inertial measurement units), can be integrated
with GPS (global position system). In fact, progresses in these years allow
the production of IMU-enabled GPS devices, that allow a GPS receiver
to work even when GPS-signals are unavailable, such as inside buildings,
tunnels, or places characterized by electronic interference. Moreover, it
is also possible to create the ManDri app for all the devices so that, after
connecting them to the gloves, dedicated accounts can be created for
chosen people allowed to use them, making easier the control, the man-
agement, and the gloves gestures customization. This device unavoidably
becomes essential and difficult to replace after trying it. These positive
premises make us understand how advantageous the choice to use these
gloves in everyday driving to improve it, make it safer, and undoubtedly
facilitate it.
Keywords: Manus VR · ManDri · IMU · GPS
1 Introduction
Before speaking in detail of gestures recognition it is essential to define an impor-

tant concept, one of the major challenge in informatics: the HCI.
Human-Computer Interaction (HCI) is a technology that has significantly
changed over the years, and has been improved day by day. It can be described
as the point of communication between the human user and the computer: an
input interface. An ideal input interface should perform well, according to some

https://doi.org/10.1007/978-3-030-15035-8_84
ManDri: A New Proposal of Manus VR Facility Integration 861
common criteria (i.e. accuracy, throughput, affordability, sociability, mobility,

usability).
Technological advances, including virtual reality, have gradually led the clas-
sical interfaces, such as mouse, keyboard, joystick and so on. Moreover, the
research on HCI is gaining momentum because of its possible applications in
areas like athletic performance analysis, surveillance, video conferencing, video
gaming, sign language recognition, lie detection, rehabilitative applications, and
so on.
Based on the criteria cited before, human gesture seems to be the most
suitable to be HCI interface and its use is increasing day by day [12,30]. We
refer to gesture as a form of non-verbal communication, including head, hand,
and body gestures. Hand Gesture Recognition has great relevance for HCI owing
to its extensive list of applications in various fields, including sign language
recognition, virtual reality, and computer games. Just think of sign language,
which is the extensively used technique when audio transmission is proscribed
or quite improbable. Since sign language is characterized by gestures as well as
the languages that are spoken, an observer should have the ability to discern
continuous sign vocabularies in real-time.
Hand postures and gestures are the common-used nonverbal communica-
tion ways, and they are the primitive skills of social intercourse for human.
They are also applied abroad on different domains, including human com-
puter/machine interaction, sign language recognition, interactive games, vision-
based augmented reality. On the other side, for communications intended for
a not audible distance, or for hearing-impaired people, the hand postures and
gestures are important tools.
Devices that sense body position and hand gestures, speech and sound, facial
expression, haptic response, and other aspects of human behavior or state, can
be used in order to make the communication between the human and computer
more natural and powerful.
To achieve natural human-computer interaction, the human hand could be
considered as an input device: indeed, hand gestures are a powerful human to
human communication modality. However, the expressiveness of hand gestures
has not been fully explored for applications. This is also due to their vastness and
variety: they change from country to country and from culture to culture; just
think that in Bulgaria the “yes” and “no” gestures are completely the opposite
to the Italian ones.
The human hand is a complex articulated object, consisting of many con-
nected parts and joints. Considering the global hand pose and each finger joint,
human hand motion has roughly 27 degree of freedom (DOF). User Interface
development requires an understanding of human hand’s anatomical structure
to determine what kind of gestures are comfortable to make [17,27]. With the
latest advances in the fields of computer vision, image processing, and pattern
recognition, real-time hand gesture classification is becoming more and more
feasible for human-computer interaction.
862 W. Balzano et al.
In the literature of gesture recognition, there are two important definitions

need to be cleared: hand posture and hand gestures. A hand posture is defined
solely by the static hand configuration and hand location without any movements
involved. A hand gesture refers to a sequence of hand postures connected by
continuous motions (global hand motion and local finger motion) over a short
time span. A hand gesture is a composite action constructed by a series of hand
postures that act as transition states. With this composite property of hand
gestures, it is natural to decouple the problem of gesture recognition into two
levels: low level posture recognition and high level gesture recognition.
Recognition of a static gesture requires only the hand shape. Once hand
shape is classified as static gesture by the trained classifier, command is given to
the computer. Unlike static gestures, dynamic ones require both the hand shape
and motion. These dynamic hand shapes are categorized into unidirectional and
multi-directional hand gestures. Unidirectional hand gestures require shape and
direction of motion of the hand, whereas multi-directional gestures require the
position of hand together with its shape [13,16,18].
Among many dynamic hand shapes, three of them are used for unidirectional
gestures, that is: turn in a direction, accelerate, decelerate. To provide them, the
Manus VR are perfect, since they manage to emulate the rotation of the hand
and every single movement of every single phalanx of each finger with extreme
precision. We will see the details of their components later.
The rest of the paper is organized as follow: Sect. 2 provides an overview of
the literature on this topic; Sect. 3 introduces ManDri; in Sect. 4, its features and
its interaction with Manus VR is explained; in Sect. 5, experimental results are
shown; finally, Sect. 6 provides conclusions and hints for future developments.
2 Related Work
VANET is a growing research field aimed to improve road safety and traffic
conditions [2,5], and path tracking techniques [3]. Various proposals have been
done to guarantee a “virtual guide”, or a connection man-car with some gestures,
but never exploiting Manus VR, meaning never with a device that guarantees
such precision. For example in [23], they presents a prototype in-car system
that allows car features (like turning lights and windows) to be controlled by
combinations of speech, gaze, and micro-gestures. They propose an interaction
concept man-car where “Micro Gestures” are directly performed with the hands
on the steering wheel, so it can be considered hands-free. For the recognition of
finger movements they use a Leap Motion controller, which is placed at the center
of the steering wheel. This is a small USB device which uses two monochrome
IR cameras at 200 Hz wide-angle (up to 150◦ ) and three infrared LEDs, that
permit to observe a certain area for a distance of about 1 m.
Similar to the Leap Motion mentioned before, authors in [14] introduce the
design method of intelligent car based on Wi-Fi video capture. The intelligent
car designed in this paper combines Wi-Fi technology and image processing
technology, and is based on Wi-Fi video capture and OpenCV gesture control
[19]. The car can carry out on-site video capture by the USB camera installed
on a router wireless, then transfers through it the collected video streaming to
the PC side, in order to achieve on-site surveillance.
The main protagonist of gesture recognition is the Kinect [15,24,29]. In [26],
for example, they have used Microsoft’s Kinect sensor to detect the human ges-
tures in a driver game. This device primarily has three sensors: IR emitter, IR
detector and RGB camera. Based on these sensors along with software develop-
ment kit (SDK), a human body is approximated using 20 body joint coordinates
in 3-dimensional space. However, for the implementation of the virtual reality
problem of car racing game, only the information on 9 joints are necessary. Using
this reduced set of sensors reduces the computation time, which is the prime con-
cern for real-time systems. The working range for Kinect sensor is from 1.2 m to
3.5 m, which makes this device perfectly acceptable for application in car racing
games, but nothing more.
Instead, in [28], they use a mm-wave radar sensor. Radar is advantageous
because it is not affected by light or sound in the environment; it (i ) can be
embedded in devices, (ii ) has very precise resolution, (iii ) offers real time recog-
nition, and (iv ) does not require recording of an image of the user. Their problem
is when they are placed too close to detectable objects, such as a gear shift. In
these cases, it often recognizes false positives, or does not recognize gestures.
In [21], they present an original contactless human machine interface for
driving car. The proposed framework is based on an image sent by a simple
camera device, which is then processed by various computer vision algorithms.
These algorithms allow the isolation of the user’s hand on the camera frame and
the translation of its movements into orders, that are sent to the computer in a
real time process. The optimization of the implemented algorithms on graphics
processing unit leads to real time interaction between the user, the computer,
and the machine. The user can easily modify or create the interfaces displayed
by the proposed framework to fit his personnel needs. Their framework works
in two steps. The first one consists in the initialization: an XML file is read,
containing the configuration to use; the second step consists in the execution of
the previous one.
In [25], they created a capacitive sensing method using electrode as sensing
input. The electrode acts as detector to retrieve the field variations at different
position of hand, in order to measure the origin of electric field distortion from the
variation in signals received due to the conductivity of the human body intercept
with capacitor itself. Then this information is used to calculate the position
of hand, to track movement of hands, and to classify gestures. This system
consists of three aluminum foil board, structured it in cube, from with each plate
acting as X, Y, Z coordinates. The cube structure is conductive so that it will
store capacitive charge with current and voltage supplied it, so the movement
of hand within the 3 plates of capacitor will act as a gesture. These gestures
will discharge the capacitor. The discharging time of each plate is different and
varies accordingly. The discharging values of each plate change according to the
situation: for instance, they very if capacitor is placed on wooden material, or on
conductive material. They have analyzed 8 gestures in two modalities (hence, 16

in total), one controlling music system, and the other controlling the accessories.
3 ManDri
The Taxonomy of gestures is important in the field of gestural interaction given

the diversity, complexity, and spontaneity of gestures in the different domains
of human-machine interaction. A well-structured taxonomy includes efficient
classes and labels, and is able to effectively distinguish the gestures in the various
application domains with the various technologies used.
In order to deduce, therefore, the intent of the gesture there is a sub-
classification: to be precise, there are five subclasses of Gesture Styles. The word
“gesture” is used for many different phenomena involving human movements,
especially of the hands and arms. Only some of these are interactive and com-
municative. Among them, we mention (i ) Deictic Gestures, that has the aim
of establishing the identity or the spatial position of an object in the context
of the application domain; (ii ) Gesticulation, that is considered as one of the
most natural forms of gesture and is commonly used in conjunction with speech
conversational interfaces; (iii ) Language Gestures, based on linguistics and per-
formed using a combination of individual signs and gestures; (vi ) Manipulative
Gestures, whose purpose is to control some entities by applying a close relation
between the real movements of the hand/arm that is handling the manipulated
entity; (v ) Semaphoric Gestures, among the most widely applied styles, that
can include static poses (e.g. index and thumb united to represent the “ok”
symbol) or dynamic movements (e.g. hand sways in the act of greeting) unlike
manipulative gestures that are mainly dynamic.
Once the types of gestures are defined, we can focus on their recognition. The
advantages of the gesture input based on data glove are that the amount of input
data is small, the speed of recognition is high, the three-dimensional information
of the hand and the motion information of the fingers can be obtained directly,
and many kinds of gestures can be identified in real time. Currently, gesture
recognition mostly uses Template Matching, Neural Networks, Hidden Markov
Model, Support Vector Machine and so on. In particular, to check the gesture
performed suggests the following recognizer algorithm:
The concept of “is closer” is widely defined by [1,20,22].
The idea, therefore, is to scroll through the list of feature vectors of gestures,
extract the gesture, check its feasibility, and then apply it to the car. To check
the feasibility of gesture a simple method with a switch with various cases could
be the solution:
As we have already said, there can be a vastness of gestures to be used as
actions associated with the car, but the basic gestures subsequently presented
are: turn left, turn right, accelerate, decelerate and change mode.
Algorithm 1. High-level algorithm: Gesture Recognizer

input : Vin : feature vector of gesture
output: Vo : gesture performed
1 for each element Vi in list of gesture do
2 if Vi is closer to Vin then
3 Vo = Vi ;
4 end
5 end
6 return Vo ;
Algorithm 2. High-level algorithm: Check Feasibility

input : Vin : feature vector of gesture
output: boolean
1 switch Vin do
2 case gesture a do
3 if gesture a is not feasible to the current state of the car then
4 return false
5 end
6 end
7 case gesture b do
8 ...
9 end
10 ...
11 endsw
12 return true;
3.1 Basic Driving Gestures
With a simple movement of the hand with the thumb raised to the left you can
decide whether to take the road on left (Fig. 1a). The same specular movement,
instead, will allow you to take the right one (Fig. 1b). With a movement of the
hand with the index raised forward it will be possible to start or accelerate the
car, depending on whether it is stationary or moving (Fig. 1c). It will be the
CheckFeasibility algorithm to manage this logic, in fact in the algorithm there
is a check on the current state of the car, and based on it, it will be possible to
understand the action that the car must perform, in this case start or accelerate.
Instead a backward movement with the thumb will make it decelerate, stop or
reverse gear (Fig. 1d). By doing the pinch with index and thumb, as shown in
Fig. 1e, it will be possible to change the command mode, moving from the driving
gestures to the gestures for the GPS.
We can imagine how many gestures are possible, and how many components
we can apply this technology: turn signals, seats, air conditioning, lights, anti-
theft, in short, to every component of the car. The usefulness of this technology
takes its maximum level in dangerous situations. In these situations, the human
Fig. 1. The turn left (a), right (b), start/accelerate (c), stop/decelerate/reverse (d),
and change mode (e) gesture.
instinct urges us to put our hands forward to protect ourselves, or make a sudden
turn, then a sudden movement, in the same way in the guide if a sudden obstacle
appears before us, like a dog, a landslide; instinct would push us to protect
us, reading and processing these gestures we could make sure to stop the car
immediately, certainly stopping it before it would stop if the driver should press
the brake pedal, saving the life of the passenger and the life of those around us.
This idea can perfectly be combined with [8].
4 ManDri Features: Manus VR
The main character of the system is Manus VR. These are gloves for high-end
touch data that have been created to capture human hands motion to use them
as a perfect input device. They are controllers capable of capturing and repro-
ducing, respectively, the position and the movements of the hands and fingers,
by means of advanced trackers and sensors. Their shape is different from a hypo-
thetical controller, because these are real technological gloves. One glove contains
two inertial measurement units, one located on the top of your hand and one
on your thumb which measures it moving separately from the other fingers.
An IMU is an electronic device that can calculate and reports an angular rate,
body’s specific force, and sometimes the magnetic field surrounding the body,
using a combination of gyroscopes and accelerometers, sometimes also magne-

tometers. Recent developments allow for the production of IMU-enabled GPS
(Global Positioning System) devices. Furthermore, [6,7,10] are valid ideas and
methods for determining the precise position of the user in space through GPS.
Thanks to the 11 sensors (2 for finger and 3 for thumb) and to the IMU system
they succeed follow the movement of the fingers and hands in real time, natural
movement can be perfectly simulated managing what the hand is doing at all
times, from the movement of every single phalanx of each finger, to the rotation
of the hand. A haptic motor above the glove can give tactile feedback to the
user. This data is sent with low latency (<5 ms) through a wireless transmitter.
In order to understand how gloves can be useful for this purpose it is advisable
to know in detail their features: each glove contains eleven sensors, a gyroscope,
an accelerometer and a magnetometer to measure the orientation of the hand.
All the hardware components are completely sealed inside the gloves, so you can
wash them like any work glove. Each finger contains two sensors that track the
movement. In addition, the thumb, a finger much more complicated to handle
than others, has an additional sensor that manages its rotation. The movements
of all the fingers are therefore perfectly traced, ensuring a unique precision. The
gloves can be washed by hand, thanks to their water-based coating, and they are
completely wireless, so an extremely addictive experience is guaranteed without
wires obstruction. The latency, response speed of the system, is just 5 ms. Each
glove contains a fully vibration motor for tactile feedback.
The validation is in terms of time needed to correctly recognize a gesture: so,

if the gesture is recognized in a time less than or equal to one second. This
verification was performed via script at run-time, which is during the execution
of a demo application. The script is thought to accept the gesture according to
the time spent to recognize it (which can be one second at most). In a preliminary
verification phase, the system is directly tested by five users. The number of tests
performed for each gesture is equivalent to 50. The test consists in repeating the
gestures consecutively and uninterruptedly. This procedure was repeated for all
Table 1. Accuracy results for the considered gestures
Test Actual gesture

Turn left Turn right Forward Backward Pinch
USER A 98% 98% 96% 96% 97%
USER B 98% 98% 97% 97% 96%
USER C 98% 98% 97% 97% 95%
USER D 98% 97% 97% 98% 96%
USER E 99% 99% 98% 97% 95%
the gestures implemented in the HGR (Hand Gesture Recognition) System [11].
Results are shown in Table 1.
As we can see, results for almost all the gestures are almost perfect: indeed,
the lowest success rate is 97%. The lowest success rate is obtained with the pinch
gesture, that is recognized less than the others.
Gesture vocabulary customization in vehicular context can help drivers reduce

distraction due to potential cognitive workload. Special care must be given in
deciding how many (and which) mappings of the gestural functions should be
customizable, with the aim to be easily memorized by drivers, without caus-
ing too much mental stress. Moreover, other technical aspects should be recog-
nized, such as machine learning algorithms contribution to create an effective
customization tool.
To make the gestures customization possible, an app for Android and IOS
devices would be very useful, with the ability to calibrate the Manus VR with the
user’s hands according to his size, to create many different profiles, maybe one for
each family member. With app, they can also check all their useful information,
such as the gloves battery, and their action history. Among the many possible
functions, we mention the one allowing the use only to the selected people,
through a password or fingerprints, or even facial recognition [4,9].
References
1. Ahmed, A., Loo, C., Obo, T.: Neuro-fuzzy model with subtractive clustering opti-
mization for arm gesture recognition by angular representation of kinect data. In:
6th ICIEV-ISCMHT, pp. 1–6. IEEE (2017)
hoc networks. In: International Conference on NBis, pp. 960–969. Springer (2018)
7. Balzano, W., Stranieri, S.: A logic range-free algorithm for localization in wireless
sensor networks. In: The 24th International DMSVIVA (2018)
8. Balzano, W., Vitale, F.: RADS: a smart road anomalies detection system using
vehicle-2-vehicle network and cluster features
9. Balzano, W., Vitale, F.: FiDGP: a smart fingerprinting radiomap refinement
method based on distance-geometry problem. In: International Conference on
NBis, pp. 970–978. Springer (2018)
10. Balzano, W., Vitalem F.: GER-EN–GNSS error reduction using an elastic network
based on V2V and LIDAR. In: International Symposium on Cyberspace Safety
and Security, pp. 124–131. Springer (2018)
11. Benmoussa, M., Mahmoudi, A.: Machine learning for hand gesture recognition
using bag-of-words. In: ISCV, pp. 1–7. IEEE (2018)
12. Bhuiyan, R., Tushar, A., Ashiquzzaman, A., Shin, J., Islam, M.: Reduction of
gesture feature dimension for improving the hand gesture recognition performance
of numerical sign language. In: 20th International Conference of ICCIT, pp. 1–6.
IEEE (2017)
13. Chiang, T., Fan, C.: 3D depth information based 2D low-complexity hand posture
and gesture recognition design for human computer interactions. In: 3rd ICCCS,
pp. 233–238. IEEE (2018)
14. Dai, G., Wang, P.: Design of intelligent car based on WiFi video capture and
OpenCV gesture control. In: CAC, pp. 4103–4107. IEEE (2017)
15. Dai, G., Yu, L., Huang, J.: Dynamic and interactive gesture recognition algorithm
based on kinect. In: CCDC, pp. 3479–3484. IEEE (2017)
16. Dehankar, A., Jain, S., Thakare, V.: Performance analysis of RTEPI method for
real time hand gesture recognition. In: ICISS, pp. 1031–1036. IEEE (2017)
17. Gajjar, V., Mavani, V., Gurnani, A.: Hand gesture real time paint tool-box:
machine learning approach. In: ICPCSI, pp. 856–860. IEEE (2017)
18. Hussain, S., Saxena, R., Han, X., Khan, J., Shin, H.: Hand gesture recognition
using deep learning. In: ISOCC, pp. 48–49. IEEE (2017)
19. Khan, M., Mishra, K., Qadeer, M.: Gesture recognition using OpenCV. In: 7th
CSNT, pp. 167–171. IEEE (2017)
20. Lei, T., Jia, X., Zhang, Y., Zhang, Y., Su, X., Liu, S.: Holoscopic 3D micro-gesture
recognition based on fast preprocessing and deep learning techniques. In: Auto-
matic Face & Gesture Recognition (FG 2018), pp. 795–801. IEEE (2018)
21. Magoulès, F., Zou, Q.: GPU accelerated contactless human machine interface for
driving car. In: 16th International Symposium on DCABES, pp. 7–10. IEEE (2017)
22. Maharani, D., Fakhrurroja, H., Machbub, C., et al.: Hand gesture recognition using
k-means clustering and support vector machine. In: ISCAIE 2018, pp. 1–6. IEEE
(2018)
23. Neßelrath, R., Moniri, M., Feld, M.: Combining speech, gaze, and micro-gestures
for the multimodal control of in-car functions. In: 12th International Conference
on IE, pp. 190–193. IEEE (2016)
24. Plouffe, G., Cretu, A.: Static and dynamic hand gesture recognition in depth data
using dynamic time warping. IEEE Trans. Instrum. Meas. 65(2), 305–316 (2016)
25. Rau, M., Agarkar, B.: 3D GRS for car infotainment. In: International Conference
on ICICET, pp. 1–4. IEEE (2018)
26. Saha, S., Lahiri, R., Konar, A., Ralescu, A., Nagar, A.: Implementation of gesture
driven virtual reality for car racing game using back propagation neural network.
In: SSCI, pp. 1–8. IEEE (2017)
27. Sasaki, A., Hashimoto, H.: High degree-of-freedom hand model driven by lower
degree-of-freedom input. In: SII, pp. 722–727. IEEE (2017)
28. Smith, K., Csech, C., Murdoch, D., Shaker, G.: Gesture recognition using mm-wave
sensor for human-car interface. IEEE Sens. Lett. 2(2), 1–4 (2018)
29. Varshini, M., Vidhyapathi, C.: Dynamic fingure gesture recognition using kinect.
In: ICACCCT, pp. 212–216. IEEE (2016)
30. Wang, K., Zhao, R., Ji, Q.: Human computer interaction with head pose, eye
gaze and body gestures. In: Automatic Face & Gesture Recognition (FG 2018),
pp. 789–789. IEEE (2018)
A Smart Compact Traffic Network Vision
Based on Wave Representation
Walter Balzano(B) , Aniello Murano, Loredana Sorrentino,

and Silvia Stranieri

walter.balzano@gmail.com, murano@na.infn.it,
loredana.sorrentino@unina.it, silviastranieri1047@gmail.com
Abstract. VANET constitutes a huge research area due to its potential

in traffic management and road safety. In this paper, we propose a novel,
smart, and compact representation of vehicular networks. Starting from
the standard graph representation, we extract a signal assigning a con-
gestion factor to each vehicle, so that highly jammed traffic areas can
be immediately detected by identifying the highest peaks of the wave.
The way the signal is built provides useful information about vehicles
distribution throughout the network, producing as result a simple but
very meaningful wave characterizing the corresponding VANET.
Keywords: VANET · Wave representation
1 Introduction
Wireless Sensor Networks (WSNs) are made of low-cost and low-power sensors
able to communicate and perform distributed tasks, often by self-organizing
into clusters. These networks can be largely employed for different purposes,
such as sensing, event detection, localization [8,9,11,21]. The WSN features of
congestion control, self-configuration, and energy awareness let their use keep
growing. In particular, congestion in WSN constitutes a concrete challenging
issue, since it determines a network performance decrease, as a consequence of
an amount of requests which is higher than the ones it can be actually satisfied.
As pointed out in [3], the increasing interest in the last years is due to the
fact that rapid deployment and fault tolerance characteristics of sensor networks
make them a promising sensing technique for military purposes. Moreover, the
sensor ability of tracking the movement of small animals, as well as monitoring
the surrounding ambient, makes them suitable for environmental aims.
An interesting application of WSN is in the vehicular context [22]: indeed, a
new emerging field is the vehicular ad hoc network, where the vehicles play the
role of sensors in a WSN [4]. Due to the increasing number of road accidents,
the need for a stable communication system between vehicles is growing. Cars
can broadcast data and information about exceptional events occurrence via
https://doi.org/10.1007/978-3-030-15035-8_85
A Smart Compact Traffic Network Vision Based on Wave Representation 871
wireless media, letting the rest of the vehicles be aware of the condition inside
the network [16]. This mechanism has a significant impact not only on safety,
but also in traffic management. Due to the importance of such a network, in this
work we provide a smart VANET representation by including in a single compact
wave a lot of useful information about node congestion, traffic, and network
clusterization. Such a wave is obtained by computing the value of congestion,
named congestion factor, of any vehicle in the network, so that highly jammed
traffic areas can be immediately detected by identify the highest peaks of the
wave. Essentially, we propose a way to obtain the highest amount of information
about the network, with as few data as possible.
Outline of the Paper
The rest of the paper is organized as follows. In Sect. 2, the state of the art
is analyzed. In Sect. 3, our novel network representation is provided. Here we
distinguish the phase in which the signal is built from the one in which the
signal is analyzed and understood. In this section we also provide some examples
of use, just for the purpose of illustration. Finally, Sect. 4 gives the conclusions
and some hints for future developments of this paper, working mainly on the
obtained wave.
2 Related Work
Vehicular ad hoc networks (VANETs) has become an active area of research

due to its potential to improve vehicle and road safety, and traffic efficiency:
as explained in [25], these kind of networks belong to Intelligent Transporta-
tion System (ITS) field, where each vehicle can receive or send information to
the network. Vehicles are equipped with on-board units in order to be able to
communicate not only between each other, but also with road infrastructure
elements.
Each car in a VANET can communicate with its adjacent nodes directly
and, via multi-hop, communication can be performed also with farther vehicles,
as pointed out in [13]. Nowadays, VANETs are a hot research topic because of
the variety of applications: in [15], they propose some of them, such as vehicle
collision warning, cooperative driving, map location, automatic parking [5], path
tracking [10], and so on.
Many researchers in the last years based their studies on VANETs issues,
such as routing protocols, and clustering algorithms: in [18], they point out
that routing is a challenge due to high dynamics of VANETs, and they analyze
a position-based routing approach in city environment; similarly, in [19], they
propose a Connectivity-Aware routing for inter-vehicular communication, by
integrating the needs of locating destination and finding paths between source
and destination. [17] provides a survey on routing protocols for VANETs.
A further hot topic in this field is clusterization, that provides a clever way to
disseminate data through the network, by improving communication and reduc-
ing redundancy. This is the reason way many authors made an effort to propose
new clustering algorithms suitable for vehicular networks. As instance, we men-

tion authors in [26], which propose a multi-hop clustering schema to establish
stable vehicle groups. Furthermore, we have already faced some VANETs issues
in our recent work, by also focusing on clusterization techniques. In particu-
lar, in [12], we provide a non-exclusive clustering approach, by modifying the
DBSCAN algorithm, by revisiting the standard communication framework with
a centralized approach. Indeed, we observed that a centralized system brigs dif-
ferent benefits, among them: (i ) the simplification of clustering process, (ii ) a
full knowledge of the network topology, and (iii ) the ability of collecting data
and providing statistics. Moreover, in [6], we propose a new clustering technique
based on two kind of cluster-head election, according to the road configuration
(one-way or two-ways). Specifically, the cluster-head is chosen as the farthest
node from the source of information. This choice allows to optimized the node
distribution in clusters, and to minimize redundancy in transmission. For further
details on VANETs clustering proposals consult [24].
Although many authors focused on VANETs aspects and challenges, as far
as we know, this is the first work addressing the VANETs representation. Even
in Google Maps, the most popular mapping app used by more than 150 millions
of unique users monthly according to a 2018 statistics reported in [2], a road
representation emphasizing the traffic congestion is not treated enough. Here,
we provide a smart and compact representation of such a network, by computing
a congestion factor for each vehicle belonging to it, aimed to map the network
to a wave on a Cartesian coordinate system. The result is a very intuitive rep-
resentation of the starting VANET, providing useful information, such as traffic
distribution and network clusterization.
3 Wave Construction over Traffic Network
In this section we propose a novel algorithm to produce an opportune wave

representation of a generated traffic network. Before focusing over the core of
our work, we provide the definitions of some terms used in this section.
3.1 Preliminaries
As follows, we provide some basic concepts used in the rest of the paper. As
these are common definitions, an expert reader can skip this part.
We consider a traffic network as a undirected graph G = (V, E) in which the
set of nodes V corresponds to the vehicles in the network, with |V | = n, and the
set of edges E corresponds to the links, if any, between the nodes, with |E| = m.
We define R as the distance coefficient, i.e. the maximum value within which
two vehicles can communicate with each other. Such a value is calculated refer-
ring to the measurement of the power present in a received radio signal, i.e. the
received signal strength indicator (RSSI). The RSSI values are measured in dBm
and have typical negative values ranging between 0 dBm (excellent signal) and
−110 dBm (extremely poor signal) [20]. In particular, it is intuitive to say that
increasing the distance, the RSSI decreases. Moreover, given two nodes u, v ∈ V,
we define d(u, v) ∈ [0, R] the euclidean distance between u and v. We can state
that an edge (u, v) ∈ E exists iff d(u, v) ≤ R, and we indicate with weight(u, v)
its weight.
Since in a vehicular network d(i, j) = d(j, i) (with i = j), we directly provide a
normalized graph, by computing the average of these two distances and by assign-
ing this value to the edge linking i and j, so that the corresponding adjacency
matrix is symmetric. Finally, for each v ∈ V define neig(v) = {v |(v, v ) ∈ E}
and degree(v) = |neig(v)| as the set of v-neighbors and its number, respectively.
3.2 Construction Phase

The idea of our algorithm is to construct a smart compact vision of traffic net-
work based on wave representation, starting from randomly generated points in
the space (based on the normal distribution). In order to obtain a signal that
reflects the congestion of any vehicle in the network regardless the way it is vis-
ited, we compute a congestion factor through a function f for each node of the
network:
f : V −→ [0, 1] (1)
This factor is parametric on R and is computed as the difference between the
ideal congestion and the local congestion. Given the node v, the ideal congestion
represents the situation where all the neighbors are at maximum distance from v:
ideal(v) = degree(v) ∗ R (2)
Instead, the local congestion is an arithmetic average over the adjacents of v:

u∈neig(v) weight(u, v)
local(v) = (3)
degree(v)
Fig. 1. Example of misleading congestion factor computation
We can observe that the congestion degree of a node does not depend only
on the number of its neighbors, but mostly on their congestion. Thus, we need
to take into account the congestion of each neighbor of the considered node, in
order to avoid the situation as the one shown in Fig. 1, in which the nodes A and
B have the same number of neighbors, but the congestion of such neighbors is
different. Indeed, in the left side we have a higher congestion level with respect
to the right side, but with the previous formula A and B would have had the
same congestion factor.
For this reason, we introduce a weighted ideal congestion:

weighted ideal(v) = (degree(u) − 1) ∗ ideal(v) (4)
u∈neig(v)
Notice that when we compute the degree of the neighbors of a given node v,
we decrease it by 1 in order not to consider v again. Once the congestion factors
are computed, they are normalized, dividing them by the maximum congestion
factor of the considered network.
The network representation we want to build is f (V ), with f as defined in
Eq. 1. A key point in the wave representation is the order in which the nodes are
placed on the x axis. Our goal is to obtain a signal that immediately highlights
the most congested areas and which points belong to them. For this reason,
we propose a cluster-oriented visit of the network that, starting from a random
point, continues the visit of neighbors, putting them in the same connected area,
as long as a certain distance is not overcome. With “connected area” we mean a
set of related nodes. It is important to notice that this is not a new clusterization
technique, but an alternative way of visiting a network by detecting areas of
nodes connected within a certain distance. To this aim, we introduce a tolerance
such that, given a node v and its neighbor u, if d(v, u) ≤ R − , u is in the same
connected area as v. It is necessary to notice that we distinguish the congestion
factor of isolated nodes by the one of nodes having only edges with weight ≥R−,
by preserving them in the network. Once the connected areas are obtained, the
nodes are placed on the x axis in such a way that the ones belonging to the same
connected area are contiguous. In particular, for each connected area we put
the elements belonging to it in increasing order with respect to the congestion
factor. This choice makes easier the detection of connected areas looking at the
signal.
3.3 Signal Rendering
In this section, we propose some examples of network transformation, by analyz-

ing all the information we can infer. The examples below are generated accord-
ing to the following parameters: the number of nodes n, maximum distance R
allowed between two nodes, a tolerance introduced in the section above, and
the standard deviation sd, used to produce random points through a normal
distribution.
Starting from the wave representation, we are able to understand the network
congestion, as well as identify the different connected areas. Indeed, according to
how the signal is built, a null congestion factor corresponds to an isolated node,
and each high-low transition induces a new connected area, but the opposite does
Fig. 2. Random generation of a network with n = 35, sd = 10, R = 10, = 0.3 (a)
and the corresponding wave representation (b)
not hold in general. Hence, this is a necessary but not a sufficient condition for
the starting of a new area. This means that there could be changes of area
hidden by the signal, when the highest congestion factor of the first connected
area is smaller than the lowest one of the second connected area, introducing
false negatives. As shown in Fig. 2(b), which is the signal corresponding to the
graph Fig. 2(a), it is intuitive to observe that the most congested nodes are 24
and 29. Another information easy to deduce is about the areas of the network
made by single nodes. Indeed, they are identified by the points whose congestion
factor coincides with the dashed line, i.e. 3, 11, 33, and 35. In order to detect
the remaining connected areas, we need to retrieve the high-low transitions,
corresponding in this network to the points 9, 29, 26, and 27. By analyzing the
network in Fig. 2(a), we would have expected a change of area between the nodes
4, 13 and 8, 17, 22, that is hidden by the signal because of false negatives. The
same happens between the nodes 22 and 7.
The signal obtained as shown in the examples, is determined by visiting the
nodes of the network starting from the one having the smallest x-coordinate.
By changing the starting point, clearly the congestion factor of each node
stays the same, but the resulting wave can be a permutation of the peaks in
the current signal. This could be a limitation for comparisons between waves.
For this reason, we introduce a canonical form of the signal, obtained by chang-
ing the order of the nodes on the x axis: not only the nodes are ordered in
increasing order of congestion factor inside each connected area, but also each
connected area is ordered in decreasing order of maximum congestion factor on
the x axis. Through this normalization, we obtain a signal having the connected
area with the highest congested node on the left side, and the single nodes on
the right side.
With this approach, false negatives are also reduced. Indeed, it is no longer
possible that the highest congested factor of a previous connected area is smaller
than the lowest factor of the next area, since they are ordered, but false negatives
can still occur, as formally reported in the following theorem:
Theorem 1. Given two successive connected areas a1 and a2 , let hi and si
(with i ∈ [1, 2]) be the highest and lowest congestion factors respectively for the
corresponding area, then:
false negative =⇒ a2 made of a single node having as congestion factor h1
Fig. 3. Random generation of a network with n = 35, sd = 10, R = 10, = 0.2 (a)
and the corresponding wave representation (b).
Fig. 4. Random generation of a network with n = 35, sd = 8, R = 10, = 0.2 (a) and
the corresponding wave representation (b).
Proof. Let us assume the premise true, thus we have a false negative. The
following inequalities hold:
⎧
⎨ h1 ≥ h2 by construction
⎪
h1 ≤ s2 by def inition of f alse negative
⎪
⎩
s2 ≤ h2 trivially
Hence, the only possibility is that s2 = h2 = h1 .

The example in Fig. 3(b) shows the signal construction following the random
generation of the corresponding network in Fig. 3(a). As expected, the peaks
decreases from left to right (with 21 the most congested vehicle), each high-low
transition indicates a change of connected area, and the nodes whose values
correspond with the dashed line are isolated vehicles (2, 10, 19, 35).
By slightly changing the standard deviation value, we generate a more con-
nected network such as the one shown in Fig. 4(a). Analyzing the corresponding
signal in Fig. 4(b), we can deduce that the most congested node is 18, there are
three isolated nodes (4, 34, 10), and high-low transitions occur correspondingly
to the nodes 18, 3, 15, 5, 28 and 34 identifying the end of a connected area.
Other examples my be obtained by modifying the input parameters: the
result is a network with a different distribution of nodes. By decreasing the
maximum distance allowed between any pair of nodes, we automatically make
the radio signal less powerful. As a consequence, the visibility between nodes is
reduced and we produce a network which is more disconnected. The correspond-

ing signal, hence, presents an higher number of isolated nodes (whose congestion
factor correspond to the zero line).
On the other hand, if we increase the tolerance, we have a higher probability
that nodes far enough do not belong to the same connected area, despite being
linked. Indeed, the greater is the tolerance, the less is the maximum distance
allowed for two nodes to be in the same connected area.
4 Conclusions and Future Works

VANET constitutes a very promising research field due to the increasing num-
ber of vehicles equipped with wireless devices. This number, according to how
estimated in [1], will have increased more than three times the 2018 value in
seven years. Vehicular environments represent challenging but fascinating sce-
nario in which we find a huge amount of applications [7]. Among the best known,
we mention the Traffic information systems [23], as well as parking techniques
development [14].
In this work, we provide a smart wave representation of a network, where
vehicles are linked to each other according to RSSI value. The obtained signal
allows to identify the traffic condition of a certain environment, and how the
vehicles form connected areas according to their distribution. This work opens
up several possible future scenarios. Indeed, we are planning to exploit this repre-
sentation to preform comparisons between the waves, corresponding to networks,
and to extract similarity measures.
References
1. Statista: IoT connected devices installed base worldwide from 2015 to 2025. www.
statista.com/statistics/471264/iot-number-of-connected-devices-worldwide/
2. Statista: Most popular mapping apps in the U.S. 2018, by audience. www.statista.
com/statistics/865413/most-popular-us-mapping-apps-ranked-by-audience/
3. Akyildiz, I.F., Su, W., Sankarasubramaniam, Y., Cayirci, E.: Wireless sensor net-
works: a survey. Comput. Netw. 38(4), 393–422 (2002)
4. Al-Sultan, S., Al-Doori, M., Al-Bayatti, A., Zedan, H.: A comprehensive survey on
vehicular ad hoc network. J. Netw. Comput. Appl. 37, 380–392 (2014)
5. Balzano, W., Barbieri, V., Riccardi, G.: Smart priority park framework based on
DDGP3. In: WAINA 2018. IEEE (2018)
7. Balzano, W., Murano, A., Vitale, F.: V2V-EN–vehicle-2-vehicle elastic network.
continuous training. In: Proceedings of WAINA 2016. IEEE Computer Society
(2016)
9. Balzano, W., Murano, A., Vitale, F.: EENET: energy efficient detection of network
changes using a wireless sensor network. In: Proceedings of CISIS 2017. Springer
(2017)
approach for cooperative traffic control systems. In: Proceedings of 3PGCIC 2016.
Springer (2016)
management. In: Proceedings of WAINA 2016. IEEE Computer Society (2016)
using V2V/V2I and DGP-class problem. In: Proceedings of WAINA 2017. IEEE
Computer Society (2017)
15. Boukerche, A., Oliveira, H., Nakamura, E.F., Loureiro, A.: Vehicular ad hoc net-
works: a new challenge for localization-based systems. Comput. Commun. 31(12),
2838–2849 (2008)
16. Cunha, F., Villas, L., Boukerche, A., Maia, G., Viana, A., Mini, R., Loureiro, A.:
Data communication in VANETs: protocols, applications and challenges. Ad Hoc
Netw. 44, 90–103 (2016)
17. Li, F., Wang, Y.: Routing in vehicular ad hoc networks: a survey. IEEE Veh.
Technol. Mag. 2(2), 12–22 (2007)
18. Lochert, C., Hartenstein, H., Tian, J., Fussler, H., Hermann, D., Mauve, M.: A
routing strategy for vehicular ad hoc networks in city environments. In: Intelligent
Vehicles Symposium. IEEE (2003)
19. Naumov, V., Gross, T.R.: Connectivity-Aware Routing (CAR) in vehicular ad-hoc
networks. In: Proceedings of INFOCOM 2007. IEEE (2007)
20. Parameswaran, A.T., Husain, M.I., Upadhyaya, S., et al.: Is RSSI a reliable param-
eter in sensor localization algorithms: an experimental study. In: Proceedings of
F2DA 2009. IEEE (2009)
21. Raghavendra, C.S., Sivalingam, K.M., Znati, T.: Wireless Sensor Networks.
Springer, London (2006)
22. Saha, A.K., Johnson, D.B.: Modeling mobility for vehicular ad hoc networks. In:
Proceedings of VANET 2004. ACM (2004)
23. Singh, A., Kumar, M., Rishi, R., Madan, D.K.: A relative study of MANET and
VANET: its applications, broadcasting approaches and challenging issues. In: Pro-
ceedings of CCSIT 2011. Springer (2011)
24. Vodopivec, S., Bešter, J., Kos, A.: A survey on clustering algorithms for vehicular
ad-hoc networks. In: Proceedings of TSP 2012. IEEE (2012)
25. Zeadally, S., Hunt, R., Chen, Y., Irwin, A., Hassan, A.: Vehicular Ad Hoc Net-
works (VANETs): status, results, and challenges. Telecommun. Syst. 50(4), 217–
241 (2012)
26. Zhang, Z., Boukerche, A., Pazzi, R.: A novel multi-hop clustering scheme for vehic-
ular ad-hoc networks. In: Proceedings of MOBIWAC 2011. ACM (2011)
Big Data Analytics for Traceability
in Food Supply Chain
Alessandra Amato(B) , Giovanni Cozzolino, and Vincenzo Moscato
University of Naples Federico II, via Claudio 21, Naples, Italy

alessandra.amato@studenti.unina.it,
{giovanni.cozzolino,vmoscato}@unina.it
Abstract. The amount of socio-economic data generated every day

has grown dramatically in recent years thanks to the widespread use
of the internet connection and the increase in the availability of elec-
tronic devices. This leads to the production of a huge amount of digital
traces of various kinds: photos, emails, call logs, information on purchases
made, financial transactions, social interactions network. Big Data are
data characterized by volume, speed and variety: they are extracted and
processed at high speed and collected in large datasets, which are made
up of data from the most varied sources and therefore not only from
structured data. Data collection is typically difficult and expensive, both
in terms of time and money; instead, the enthusiasm that surrounds
Big Data is due precisely to the perception of great ease and speed of
access to a large amount of data at low cost. Thence, in this work we
show the application of a system architecture aiming to use of Big Data
technologies for traceability in food supply chain domain.
1 Introduction
More and more, our daily actions leave behind us a digital trail: our consumption
habits, our web searches, videos uploaded or viewed online, opinions shared in
social networks, call logs, even car or our own bodies movements. In fact, the
use of any electronic devices generates information, both direct and indirect, on
people and their interactions that remain recorded and can go to create huge-
sized datasets; about two and a half million (2,500, 000, 000, 000, 000, 000) of
bytes created every day in the world are estimated. This huge availability of data
in real time, for those interested in studying human behaviour, is comparable to
a gold mine to be exploited in some way.
With the current availability of data, it is possible to study various phe-
nomena with a considerably greater amount of information to be analysed and
available at lower costs and times [1,2]. With Big Data we can not only use large
databases to validate a hypothesis already formulated, but also give voice to the
data themselves, making us not only enlighten them, but also get inspired by
them.
In this work we show an application and some examples of the use of Big
Data technologies for traceability in food supply chain domain. The paper is
https://doi.org/10.1007/978-3-030-15035-8_86
Big Data Analytics for Traceability in Food Supply Chain 881
structured as follows: Sect. 2 describes the Data model and the architecture of
our system; while Sect. 3 remarks the conclusions and the goals achieved.
2 System Description
2.1 Data Model
The data model relies on the concept of item FI that is in turn related to a given
food item, it is inspired by the data model reported in [3]. In a simplified way, we
consider an ontology O = (V, E) as a network of concepts belonging to a given
domain of interest, where a node v ∈ V represents a “concept” and an edge
e ∈ E represents a relationship between two concepts. We define an annotation
schema and a semantic annotation [4] for a food item as follows.
Definition 2.1 (Annotation Schema). Given a set of ontologies O, an Anno-

tation Schema is a particular tuple λO (A1 , ..., An , B1 , ..., Bm ), where A1 , ..., An
are attributes for which ∀i ∈ [1, n], ∃ O = (V, E) ∈= s.t. dom(Ai ) ⊆ V (i.e.,
Ai assumes values corresponding to nodes of some ontology), and B1 , ..., Bm are
attributes for which ∀j ∈ [1, m], ∃ O = (V, E) ∈ O s.t. dom(Bj ) ⊆ V (i.e., Bj
does not assume values corresponding to nodes of any ontology).
In other words, the attributes A1 , ..., An are ontological attributes and corre-
spond to concepts that are relevant for the specific domain(s) being modelled. In
turn, non-ontological attributes B1 , ..., Bm can contain other useful information,
such as measures returned by the sensors attached to food items, or multime-
dia objects (e.g., audio, video, images, text, and 3D models) characterized by
low-level features and metadata. In particular, we can adopt both “literals” or a
set of Uniform Resource Identifiers (URIs) that allow access to related informa-
tion according to the Linked Data/Linked Open Data (LD/LOD) paradigms, as
values of the annotation attributes.
Definition 2.2 (Semantic Annotation). Given a set of ontologies O, an

annotation schema λO , and FI, a Semantic Annotation of FI is a tuple λO (F I) =
(a1 , ..., an , b1 , ..., bm ), where ∀i ∈ [1, n], ai ∈ dom(Ai ) and ∀j ∈ [1, m], bi ∈
dom(Bi ).
Using various sets of ontologies/taxonomies and semantic annotations, we

can thus describe a food item from different points of view supporting several
applications. A large set of relationships can also be instantiated among food
items, and the entire Knowledge Base (KB) can be modeled as a graph.
Definition 2.3 (Food Traceability Knowledge Base). The Food Traceabil-

ity Knowledge Base is a graph G = (C, R): Each node c ∈ C can be a food
item or an ontological attribute, while each edge r ∈ R represents a relationship
derived from a semantic annotation or established between two food items. Thus,
these nodes and edges are abstract data types with a set of properties (attributes
and methods).
882 A. Amato et al.
Leveraging different annotation schemas and available ontologies, our model

allows achieving interoperability goals. In fact, the Knowledge Base content can
be easily exported in the most-used formats, e.g., XML (eXtensible Markup
Language), RDF (Resource Description Framework), OWL (Ontology Web Lan-
guage), and according to the most diffused harvesting standards for food trace-
ability applications. On the other hand, the LD/LOD paradigm permits us to
deal with several problems related to data consistency and copyright constraints
in an effective manner: some items descriptions are accessible only using URI,
and thus the data management issues are connected to the related source.
2.2 Architecture
Figure 1 describes at a glance a functional overview of the proposed system in

terms of its main components. It is worth noticing that it presents the layered
architecture typical of a Big Data system [5], exploiting a related stack of tech-
nologies and it is inspired by [3].
In the data source layer, each data source is properly wrapped in order to
extract the information of interest that is then represented as required by the
data model. In particular, each Wrapper is specialized for a particular kind of
source (i.e., Sensor Networks, Social Networks, Digital Repositories, Users Data,
and Web Data Services) and must address all the interoperability issues, pro-
viding a set of functionalities to access data sources and gather all the desired
data, possibly leveraging the available APIs. Data integration problems for het-
erogeneous data sources are addressed by means of classical schema mapping
techniques or other type of approaches, according to the specific data source
[6–8]. In the data management layer, data are stored in the KB in compliance

Big Data Analytics for Traceability in Food Supply Chain 883
with the above-described data model and managed also exploiting the LD/LOD
paradigm. In addition, specific semantics to be attached to the data is provided
using the annotation schemes, including ontologies, vocabularies, taxonomies,
and so on, related to the food traceability domain. The Knowledge Base lever-
ages different advanced data management technologies (e.g., Distributed File
Systems, NoSQL (non relational SQL), and relational systems) and provides a
set of basic APIs to read/write data.
The data management layer is further enriched with two additional com-
ponents: a Geographic Information System (GIS) and a Context Management
System (CMS) [9–11]. The GIS manages the maps of the environments offering
a set of primitives to capture, store, manipulate, retrieve, analyse, and present
all types of spatial or geographical data; the CMS performs data tailoring by
exploiting context information and generates a set of geo-referenced Contextual
Areas, possibly enabling a set of services on such data views [12–15]. As a basis
for the application layer, our system provides a set of services that can be invoked
by user applications:
• Recommendation Services: to suggest personalized paths to the items and

also to help them browse multimedia collections related to a given FI;
• Information Retrieval Services: to search information of interest using
content-based facilities;
• Data Analytics Services: to obtain useful statistics about a FI or as support
for data-mining and pattern recognition applications;
• Social Comment Services: to post comments and feedback about users’ sen-
timent on social networks [16];
• Utility Services: for example, to inform users on packaging or information
about storing related to a given food item.
3 Conclusions
In this article, we showed a model and a system architecture finalized to use Big
Data technologies for traceability in food supply chain domain. In particular, we
described a scalable prototype for the management and context-driven browsing
of food production environments. The system is characterized by several features
that are typical of the modern big data systems and can be used to monitor data
about food supply chain.
References
1. Cilardo, A., Durante, P., Lofiego, C., Mazzeo, A.: Early prediction of hardware
complexity in HLL-to-HDL translation. In: 2010 International Conference on Field
Programmable Logic and Applications (FPL), pp. 483–488. IEEE (2010)
884 A. Amato et al.
3. Amato, F., Moscato, V., Picariello, A., Colace, F., De Santo, M., Schreiber, F.A.,
Tanca, L.: Big data meets digital cultural heritage: design and implementation
of scrabs, a smart context-aware browsing assistant for cultural environments.
J. Comput. Cult. Herit. (JOCCH) 10(1), 6 (2017)
4. Amato, F., Mazzeo, A., Penta, A., Picariello, A.: Building RDF ontologies from
semi-structured legal documents. In: 2008 International Conference on Complex,
Intelligent and Software Intensive Systems, CISIS 2008, pp. 997–1002. IEEE (2008)
5. Philip Chen, C.L., Zhang, C.-Y.: Data-intensive applications, challenges, tech-
niques and technologies: a survey on big data. Inf. Sci. 275, 314–347 (2014)
6. Cugola, G., Margara, A.: Complex event processing with T-REX. J. Syst. Softw.
85(8), 1709–1728 (2012)
7. Panigati, E., Schreiber, F.A., Zaniolo, C.: Data streams and data stream manage-
ment systems and languages. In: Colace, F., De Santo, M., Moscato, V., Picariello,
A., Schreiber, F., Tanca, L. (eds.) Data Management in Pervasive Systems, pp.
93–111. Springer, Cham (2015)
8. Marelli, M., Fortunato, M., Camplani, R., Schreiber, F.A., Rota, G.: Perla: a lan-
guage and middleware architecture for data management and integration in per-
vasive information systems. IEEE Trans. Software Eng. 38, 478–496 (2012)
9. Colace, F., De Santo, M., Greco, L., Moscato, V., Picariello, A.: Probabilistic
approaches for sentiment analysis: latent dirichlet allocation for ontology building
and sentiment extraction. In: Pedrycz, W., Chen, S.M. (eds.) Sentiment Analysis
and Ontology Engineering, pp. 75–91. Springer, Cham (2016)
10. Bolchini, C., Quintarelli, E., Tanca, L.: Carve: context-aware automatic view def-
inition over relational databases. Inf. Syst. 38(1), 45–67 (2013)
11. Dong, X.L., Srivastava, D.: Big data integration. In: 2013 IEEE 29th International
Conference on Data Engineering (ICDE), pp. 1245–1248. IEEE (2013)
12. Amato, F., Castiglione, A., Moscato, V., Picariello, A., Sperlı́, G.: Multimedia
summarization using social media content. Multimed. Tools Appl. pp. 1–25 (2018)
13. Amato, F., Castiglione, A., De Santo, A., Moscato, V., Picariello, A., Persia,
F., Sperlı́, G.: Recognizing human behaviours in online social networks. Comput.
Secur. 74, 355–370 (2018)
14. Amato, F., Moscato, V., Picariello, A., Sperlı́, G.: Extreme events management
using multimedia social networks. Future Gener. Comput. Syst. 94, 444–452 (2019)
15. Bartolini, I., Patella, M.: Multimedia queries in digital libraries. In: Colace, F.,
De Santo, M., Moscato, V., Picariello, A., Schreiber, F., Tanca, L. (eds.) Data
Management in Pervasive Systems, pp. 311–325. Springer, Cham (2015)
16. Colace, F., De Santo, M., Greco, L., Moscato, V., Picariello, A.: A collabora-
tive user-centered framework for recommending items in online social networks.
Comput. Hum. Behav. 51, 694–704 (2015)
The 2nd International Workshop on
Internet of Everything and Machine
Learning Applications (IOEMLA-2019)
Social Credibility Incorporating
Semantic Analysis and Machine Learning:
A Survey of the State-of-the-Art and Future
Research Directions
Bilal Abu-Salih1,2(&), Bushra Bremie1, Pornpit Wongthongtham1,

Kevin Duan1, Tomayess Issa1, Kit Yan Chan1,
Mohammad Alhabashneh1, Teshreen Albtoush1,
Sulaiman Alqahtani1, Abdullah Alqahtani1, Muteeb Alahmari1,
Naser Alshareef1, and Abdulaziz Albahlal1
1
Curtin University, Perth, Australia
{bilal.abusalih,bushra.bremie,
ponnie.clark,yuchao.duan,
tomayess.issa,kit.chan,
mohammad.alhabashneh,
Abdulaziz.albahlal}@curtin.edu.au
2
The University of Jordan, Amman, Jordan
Abstract. The wealth of Social Big Data (SBD) represents a unique opportu-
nity for organisations to obtain the excessive use of such data abundance to
increase their revenues. Hence, there is an imperative need to capture, load,
store, process, analyse, transform, interpret, and visualise such manifold social
datasets to develop meaningful insights that are specific to an application’s
domain. This paper lays the theoretical background by introducing the state-of-
the-art literature review of the research topic. This is associated with a critical
evaluation of the current approaches, and fortified with certain recommendations
indicated to bridge the research gap.
1 Introduction
The dramatic increase in the social data impact - a testimony to our growing digital
lifestyles - has taken on industries and activities ranging from marketing and adver-
tising to intelligence gathering and political influence. Interpreting such a Big Data
island is momentous to bring new perspectives and to improve business practices, yet
this revolution is still in its infancy. It is easy to assume that social data revolution is
about quantity, but social data is more than just volume. In fact, the extents of this
revolution are more wide-spreading: it is about building data infrastructures that are
needed to effectively digest the breeding of social data to infer the hoped-for added
value. This has motivated the research communities to dig deep, provide solutions and
to implement platforms into potential usage of these datasets to benefit several appli-
cations [1–7].

https://doi.org/10.1007/978-3-030-15035-8_87
888 B. Abu-Salih et al.
This paper presents a thorough survey of the state-of-the-art approaches drawn

from academic sources with the relevance to the topic of this paper. The purpose of the
review is threefold: Social Trust, Semantic Analysis, and Data Classification in the era
of SBD.
2 SBD Incorporating Trust
In modern enterprises, social networks are used as part of the infrastructure for a
number of emerging applications such as recommendation and reputation systems. In
such applications, trust is one of the most important factors in decision making.
Sherchan et al. [8] defined Trust as the measurement of confidence where a group of
individuals or communities behave predictably.
2.1 Generic-Based Trustworthiness Approaches

Generic-based trust approaches in Online Social Networks are those frameworks,
techniques, and tools developed to calculate and infer the trustworthiness values of
users and/or their content with no consideration to the domain(s) of interest which can
be extracted from the user level or post level. The trustworthiness of social media data
is now a crucial consideration. With such a vast volume of data being interchanged
within the social media ecosystems, data credibility is a vital issue, especially regarding
personal data [9].
Graph-Based Social Trust: Podobnik et al. [10] propose a model that calculates trust
between friends in a network graph based on weights of the edges between user’s
connected friends in Facebook. Agarwal and Bin [11] suggest a methodology to
measure the trustworthiness of a social media user by using a heterogeneous graph in
which each actor in the Twitter domain was presented as a vertex type in the graph. The
level of trustworthiness was measured using a regressive spread process. On the other
hand, the paper, neglects to consider the importance of weighting scheme and time
factor. Each edge category should be assessed at different credibility levels; hence, a
weighting scheme should be used. Trustworthiness values differ over time; conse-
quently, the temporal/time factor should be integrated.
Trust for Recommendation Systems: Massa et al. [12], show a web of trust as an
alternative to the standard way of ranking a user, i.e. standard recommendation sys-
tems. Further, Gupta et al. [13] present the “WTF: Who To Follow” service which is
being used as a recommendation system for the Twitter social network. This service is
used mainly as a recommendation driver and has a significant impact; by using it,
numerous new connections have been created. Further work by Gallege et al. [14] and
Sun et al. [15] propose trusted-based recommendation techniques.
Trust Incorporating Sentiment Analysis: The use of sentiment analysis techniques to
analyse the content of OSNs has significantly influenced several aspects of research. In the
context of social trust, Alahmadi et al. [16] propose a recommendation system framework
incorporating implicit trust between users and their emotions. AlRubaian et al. [17]
Social Credibility Incorporating Semantic Analysis and Machine Learning 889
present a multi-stage credibility framework for the assessment of microbloggers’ content.

The development of sentiment-based trustworthiness approaches for OSNs is discussed
further in [18–20].
2.2 Domain-Based/Topic-Specific Trustworthiness Approaches

Adding a user-domain dimension when calculating trust in social media is an important
factor, to enhance understanding of users’ interest. In this context, the notion of
domain-based trust for the data extracted from the unstructured content (such as social
media data) is significant. This is through calculating trustworthiness values which
correspond to a particular user in a particular domain. The literature of trust in social
media shows a lack of approaches for measuring domain-based trust. Several reviews
have carried out to highlight the importance of conducting a fine-grained trustwor-
thiness analysis in the context of SBD [8, 21–23]. In particular, measuring the user’s
trustworthiness in each domain of knowledge is vital to better understanding users’
behaviours in the OSNs.
3 SBD Incorporating Semantic Analysis, and Machine

Learning for Data Classification and Topic Distillation
In the latter part of the 20th century, researchers in the field of Artificial Intelligence
(AI) have become active in computational modelling of statistical analysis means
alongside with defining ontologies that would deliver automated reasoning capabilities.
This section aims to review the existing approaches into two main categories:
(i) techniques incorporating semantic analysis for domain discovery; (ii) machine
learning statistical techniques for data classification topics distillation.
3.1 SBD Incorporating Semantic Analysis

The Semantic Web (SW) was introduced by Berners Lee who provided a new vision
for the next web where data is given semantic meanings via data annotation and
manipulation in a machine-readable format [24]. Ontology-based on Gruber’s [25]
definition is the formal explicit specification of a shared conceptualization within a
domain, as a form of concepts and relationships between these concepts which is used
to describe the domain. By incorporating semantic analysis, semantic data can be
inferred from social media data. The use of ontology in the social media has been
applied widely to infer semantic data in a broad range of applications. De Nart et al.
[26] propose a content-based approach to extract the main topics from the tweets. This
approach is an attempt to understand the research communities’ activities and their
emerging trends. Chianese et al. [27] propose a data-driven and ontology-based
approach to identify cultural heritage key performance indicators as expressed by social

network users. This approach can be used in different domains but is only relevant or
ad-hoc to user domains. Michelson and Macskassy [28] use the DBpedia knowledge
base to annotate entities in users’ tweets, and extract the users’ main interests by using
the categories proposed in Wikipedia. Wikipedia as a knowledge base repository has
been utilised for topic discovery in [29, 30].
3.2 SBD Incorporating Machine Learning for Data Classification

and Topic Distillation
Machine Learning applications enable real-time predictions leveraging high quality and
well-proven learning algorithms. Based on the current dominant position and high
impact on business in several use cases, according to Gartner’s recent report on
emerging technologies1, incorporating machine learning, in particular, enhances the
decision-making process and provides valuable insights from large-scale data. Topic
distillation (a.k.a topic discovery, or topic modelling, or latent topic modelling or
statistical topic modelling) is an automatic approach to distil topics from a corpus of
words embodied in a set of documents incorporating statistical techniques [31–33].
These statistical-based techniques have been used as another means of topic
modelling and discovery in social data mining. Examples of such statistical-based
techniques that have been used are LDA (Latent Dirichlet Allocation) [34], Latent
Semantic Analysis (LSA), and recently Fuzzy Latent Semantic Analysis (FLSA) [35].
LDA is based on an unsupervised learning model to identify topics from the distri-
bution of words. In LSA, an early topic modelling method has been extended to pLSA
[36], which generates the semantic relationships based on a word-document co-
occurrence matrix. FLSA supposes that the list of documents and their embodied words
can be fuzzy clustered where each cluster can be represented by a certain topic. LDA
and similar unsupervised techniques have been widely used in several modelling
applications [37–42].
4 Critical Synthesis Review of the Current Approaches
The previous sections present an integrated review of the existing methods and
approaches in the era of SBD incorporating trust, semantic analysis and machine
learning.
• Lack of advanced domain-based trustworthiness approaches,
• Lack of managing and extracting high-level domains from the textual content of
SBD,
• Lack of domain-based techniques for dual classification.
1
http://www.gartner.com/document/3383817?ref=solrAll&refval=175496307&qid=
34ddf525422cc71383ee22c858f2238a, Visited in 25/10/2016.
4.1 Lack of an Advanced Domain-Based Trustworthiness Approaches

A Need for Domain-Based Trustworthiness: There have been some efforts apply
generic-based credibility evaluation approaches for users and their content in OSNs
[43–45]. However, they do not take the topic or subject factor into consideration; the
classification has been computed in general. Users will have a certain reputation in one
domain, but that does not always apply to any other domain. The user’s credibility
should be domain-driven. For example, evaluating users’ trustworthiness in a specific
domain has been driven by its implication in several applications such as personalized
recommendation systems [44], opinion analysis [46], expertise retrieval [47], and
computational advertising [48].
Lack of Incorporating Temporal Factor: Subsequent studies have focused on the
users’ topic(s) of interest or their domains of knowledge. However, studying users’
behaviour over time has been neglected [13, 39, 49]. The users’ behaviours may change
over time. It follows that their trustworthiness values vary over time; hence, the temporal
factor should be assimilated. Moreover, Spammers’ behaviours are unsteady as they are
not legitimate users although they pretend to be. Therefore, their “temporal patterns of
tweeting may vary with frequency, volume, and distribution over time” [50].
Lack of a sentiment analysis of conversations: In the context of social trust,
frameworks have been developed to analyse the users’ content, taking into consider-
ation the overall feelings regarding what they have chosen to expose their content
[14, 17, 19]. However, these efforts did not attempt a sentiment analysis of a post’s
replies by measuring the trustworthiness values. Sentiments of user’s followers are
significant to understand the followers’ opinions toward the user. Consequently, users
who obtain high number positive replies should attain a better reputation than users
obtain a large number of negative replies to their content.
Lack of Addressing Key Features of BD: BD technology for data storage and analysis
provides advanced technical capabilities to the process of analysing massive and
extensive data to achieve deep insights in an efficient and scalable manner. Manyika
et al. [51] listed some of the Big Data technologies such as Big Table, Cassandra (Open
Source DBMS), Cloud Computing, Hadoop (Open Source framework for processing
large sets of data), etc. Chen et al. [52] discussed the various open issues and challenges
of BD and listed the key technologies of BD. The incorporation of BD technology to
facilitate the trustworthiness measuring and inferring tools is unavoidable, especially
regarding the nature of the contents of social media which span to a huge quantum.
This has interestingly attracted researchers of social trust to leverage the BD techniques
to benefit their conducted experiments [53, 54]. However, there has been a lack of
addressing the key features of BD such as volume (i.e. massive social data datasets),
veracity (i.e. reputation of the data sources), and value (outcome product of the data
analytics). Hence, starting from the characteristics of BD and sorting out issues related
to these dimensions will be the most efficient way to address BD as well as to benefit
trustworthiness analysis endeavors with the expected outcomes of SBD Analysis.
4.2 Lack of Managing and Extracting High-level Domains

from the Textual Content of SBD
The prevailing explorations aiming to improve the understanding of the contextual
content of social data is noteworthy. Most of the existing approaches to topic distil-
lations rely on bag-of- words techniques such as LDA [33]. However, despite the
importance and popularity of these techniques to infer the users’ topics of interest,
when it comes to context on OSNs there are three main shortcomings; (i) inability to
consider the semantic relationships of terms in the user’s textual content; (ii) inade-
quacy to apply its topic modelling technique into short text messages such as tweets;
(iii) the high-level topics classifications that use these bag-of-words statistical tech-
niques are inadequate and inferior [28].
LDA extracts latent topics by presenting each topic as a words distribution. This
statistical mechanism does not consider the semantic relationships of terms in a doc-
ument [28]. Furthermore, the high-level topics classifications that use these bag-of-
words statistical techniques are inadequate and inferior. Furthermore, using this tech-
nique is inappropriate to cluster and search for users based on high-level topics [28].
On the other hand, incorporating semantic web consolidated tools such as Alchem-
yAPI™ offers a comprehensive list of taxonomies divided into hierarchies where the
high-level taxonomy represents the high-level domain and the deeper-level taxonomy
provides a fine-grain domain analysis. For instance, “art and entertainment” is con-
sidered a high-level taxonomy in which “graphic design” is one of its deep-level
taxonomy. LDA is unable to provide high-level topics such as “art and entertainment”
from a corpus of posts or tweets unless this term exists in the corpus. Semantic
analysis, conversely, extracts semantic concepts and infers high-level domains through
analysing the semantic hierarchy of each topic leveraging an ontology, which is not
possible using LDA technique.
4.3 Lack of Domain-Based Techniques for Dual Classification

This section discusses the mechanisms and limitations of the current approaches to
domain-based classification.
Domain-Based Classification (Inclusion of Both User and Post Levels): OSNs have
spurred researchers to develop several methods for discovering the main interest(s) of
their users. Due to ambiguity, shortness and nosiness of posts such as tweets [28], these
endeavours are still immature; Hence, extensive research in this area is vital [55]. For
example, Twitter tools [8, 56] are focused on the exploration of user networks to obtain
information for user interests and topics. These approaches only extract keywords to
obtain a summary of Twitter data. However, the use of keywords only cannot fully
cover user domains and may generate misleading user information. There should be
incorporation of both user level and tweet level which involve semantics of words and
accurate disambiguation for social networks study. The accurate classification of the
users’ interest assists in providing an accurate understanding of short textual content of
future tweets. This benefits several applications the aim of which is to obtain correct
domain-based trustworthiness of users and their content in OSNs.
Incorporation of Domain Ontology and Semantic Web, and Machine Learning: As

indicated earlier, the high-level topics classifications that use the bag-of-words statis-
tical techniques are inadequate, and the brevity and ambiguity of short texts make more
difficult the process of topic modelling using these statistical models [57]. Besides,
these methods do not consider the temporal factor. In other words, the users’ knowl-
edge evolves and their interest might be diverted elsewhere depending on their expe-
rience, work, study, etc. Hence, it is important to scrutinise users’ interest over time to
infer intrinsic topics of interest to users in OSNs. Hence, there is a need for a com-
prehensive approach which leverages the external domain-based Ontology and
semantic web knowledge bases to aid mitigate the disambiguation in the textual content
at the user and tweet levels. This approach should be fortified with harnessing advanced
and state-of-the-art machine learning techniques to perform domain-based classification
at the user and tweet levels.
5 Conclusion
Currently, several research efforts are undertaken to handle and manage the large scale
of SBD in the quest for added value. A thread of these efforts attempts to provide
technical solutions to cope with the volume and speed of social data generation. This is
through facilitating the process of data capturing, acquisition and storing. Another host
of scholars aim to develop data analytics solutions to enhance the quality of the collected
social data, to understand the users’ topics of interest, to classify and categorise users
into segments, and to extract better insights, thus improving the decision-making pro-
cess. Despite the significance of these endeavours and initiatives to change the per-
ception of the extracted social data, these current frameworks for SBD analysis only
partially consider SBD features. Furthermore, existing efforts are conducted merely to
address one or few issues of SBD. A comprehensive framework is required to resolve
the issues of data quality, extract hidden knowledge, and infer the credibility of users
and their content of the OSNs through extracting the domains of interest at the user and
the post level. This consolidates the development of data classification techniques which
leads to better anticipating of users’ interest in their future published content.
In this paper, an extensive survey of the existing literature is conducted. The
discussion of several approaches to measuring the generic-based and domain-based
trustworthiness in SBD is carried out followed by addressing the core literature in the
areas of semantic analysis, machine learning and data classification within the context
of SBD. The carried out efforts within the cordon of these themes are critically
reviewed which points out to the considerable achievements have been made in SBD
analytics. However, the current approaches and techniques are still insufficient in terms
of (i) the lack of domain-based trustworthiness approaches; (ii) the lack of managing
and extracting high-level domains from the textual content of SBD; (iii) the lack of
domain-based approaches for dual classification of the textual content of SBD.
References
1. Abu-Salih, B., Wongthongtham, P., Kit, C.Y.: Twitter mining for ontology-based domain
discovery incorporating machine learning. J. Knowl. Manag. 22(5), 949–981 (2018)
2. Nabipourshiri, R., Abu-Salih, B., Wongthongtham, P.: Tree-based classification to users’
trustworthiness in OSNs In: Proceedings of the 2018 10th International Conference on
Computer and Automation Engineering, pp. 190–194. ACM, Brisbane, Australia (2018)
3. Wongthongtham, P., Abu Salih, B.: Ontology and trust based data warehouse in new
generation of business intelligence: state-of-the-art, challenges, and opportunities. In: 2015
IEEE 13th International Conference on Industrial Informatics (INDIN). IEEE (2015)
4. Abu-Salih, B., et al.: Towards a methodology for social business intelligence in the era of big
social data incorporating trust and semantic analysis. In: Second International Conference on
Advanced Data and Information Engineering (DaEng-2015). Springer, Bali, Indonesia
(2015)
5. Abu-Salih, B., et al.: CredSaT: credibility ranking of users in big social data incorporating
semantic analysis and temporal factor. J. Inform. Sci. 0(0), 0165551518790424 (2018)
6. Abu-Salih, B.: Trustworthiness in Social Big Data Incorporating Semantic Analysis,
Machine Learning and Distributed Data Processing. Curtin University (2018)
7. Chan, K.Y., et al.: Affective design using machine learning: a survey and its prospect of
conjoining big data. Int. J. Comput. Integr. Manuf. 1–19 (2018)
8. Sherchan, W., Nepal, S., Paris, C.: A survey of trust in social networks. ACM Comput. Surv.
45(4), 47 (2013)
9. Passant, A., et al.: Enabling trust and privacy on the social web. In: W3C Workshop on the
Future of Social Networking (2009)
10. Podobnik, V., et al.: How to calculate trust between social network users? In: 2012 20th
International Conference on Software, Telecommunications and Computer Networks
(SoftCOM). IEEE (2012)
11. Agarwal, M., Bin, Z.: Detecting malicious activities using backward propagation of
trustworthiness over heterogeneous social graph. In: 2013 IEEE/WIC/ACM International
Joint Conferences on Web Intelligence (WI) and Intelligent Agent Technologies
(IAT) (2013)
12. Massa, P., Bhattacharjee, B.: Using trust in recommender systems: an experimental analysis,
in trust management. In: Jensen, C., Poslad, S., Dimitrakos, T. (eds.) pp. 221–235. Springer,
Berlin, Heidelberg. (2004)
13. Gupta, P., et al.: WTF: the who to follow service at Twitter. In: Proceedings of the 22nd
International Conference on World Wide Web, pp. 505–514. International World Wide Web
Conferences Steering Committee, Rio de Janeiro, Brazil (2003)
14. Gallege, L.S., et al.: Towards trust-based recommender systems for online software services.
In: Proceedings of the 9th Annual Cyber and Information Security Research Conference,
pp. 61–64. ACM, Oak Ridge, Tennessee (2014)
15. Sun, G., et al.: A social trust path recommendation system in contextual online social
networks. In: Chen, L., et al. (eds.) Web Technologies and Applications, pp. 652–656.
16. Alahmadi, D.H., Zeng, X.J.: ISTS: implicit social trust and sentiment based approach to
recommender systems. Expert Syst. Appl. 42(22), 8840–8849 (2015)
17. AlRubaian, M., et al.: A multistage credibility analysis model for microblogs. In:
Proceedings of the 2015 IEEE/ACM International Conference on Advances in Social
Networks Analysis and Mining 2015. ACM (2015)
18. Zhang, B., et al.: A trust-based sentiment delivering calculation method in microblog. Int.
J. Serv. Technol. Manag. 21(4–6), 185–198 (2015)
19. Bae, Y., Lee, H.: Sentiment analysis of twitter audiences: measuring the positive or negative
influence of popular Twitterers. J. Am. Soc. Inform. Sci. Technol. 63(12), 2521–2535 (2012)
20. Kawabe, T., et al.: Tweet credibility analysis evaluation by improving sentiment dictionary.
In: 2015 IEEE Congress on Evolutionary Computation (CEC). IEEE (2015)
21. Momeni, E., Cardie, C., Diakopoulos, N.: A survey on assessment and ranking methodolo-
gies for user-generated content on the web. ACM Comput. Surv. (CSUR) 48(3), 41 (2016)
22. Amalanathan, A., Anouncia, S.M.: A review on user influence ranking factors in social
networks. Int. J. Web Based Commun. 12(1), 74–83 (2016)
23. Ruan, Y., Durresi, A.: A survey of trust management systems for online social communities–
trust modeling, trust inference and attacks. Knowl.-Based Syst. 106, 150–163 (2016)
24. Berners-Lee, T., Hendler, J.: Publishing on the semantic web. Nature 410(6832),
1023 (2001)
25. Gruber, T.R.: Toward principles for the design of ontologies used for knowledge sharing?
Int. J. Hum Comput Stud. 43(5), 907–928 (1995)
26. De Nart, D., et al.: A content-based approach to social network analysis: a case study on
research communities, in digital libraries on the move. In: Calvanese, D., De Nart, D., Tasso,
C. (eds.) 11th Italian Research Conference on Digital Libraries, IRCDL 2015, Bolzano, Italy,
29–30 January 2015, pp. 142–154. Springer, Cham, 2016, Revised Selected Papers
27. Chianese, A., Marulli, F., Piccialli, F.: Cultural heritage and social pulse: a semantic
approach for CH sensitivity discovery in social media data. In: 2016 IEEE Tenth
International Conference on Semantic Computing (ICSC) (2016)
28. Michelson, M., Macskassy, S.A.: Discovering users; topics of interest on twitter: a first look.
In: Proceedings of the Fourth Workshop on Analytics for Noisy Unstructured Text Data.
ACM (2010)
29. Schonhofen, P.: Identifying document topics using the wikipedia category network. In:
Proceedings of the 2006 IEEE/WIC/ACM International Conference on Web Intelligence,
pp. 456–462. IEEE Computer Society (2006)
30. Hassan, M.M., Karray, F., Kamel, M.S.: Automatic document topic identification using
wikipedia hierarchical ontology. In: 2012 11th International Conference on Information
Science, Signal Processing and their Applications (ISSPA) (2012)
31. Anthes, G.: Topic models Vs unstructured data. Commun. ACM 53(12), 16–18 (2010)
32. Wang, C., et al. Markov topic models. In: Artificial Intelligence and Statistics (2009)
33. Blei, D.M., Ng, A.Y., Jordan, M.I.: Latent Dirichlet allocation. J. Mach. Learn. Res. 3(4–5),
993–1022 (2003)
34. Blei, D.M., Ng, A.Y., Jordan, M.I.: Latent Dirichlet allocation. J. Mach. Learn. Res. 3(Jan),
993–1022 (2003)
35. Karami, A., et al.: Fuzzy approach topic discovery in health and medical corpora. Int.
J. Fuzzy Syst. 1–12 (2017)
36. Hofmann, T.: Probabilistic latent semantic indexing. In: Proceedings of the 22nd Annual
International ACM SIGIR Conference on Research and Development in Information
Retrieval. ACM (1999)
37. Chen, Y., et al.: Topic modeling for evaluating students; reflective writing: a case study of
pre-service teachers’ journals. In: Proceedings of the Sixth International Conference on
Learning Analytics & Knowledge. ACM (2016)
38. Nichols, L.G.: A topic model approach to measuring interdisciplinarity at the national
science foundation. Scientometrics 100(3), 741–754 (2014)
39. Weng, J., et al.: Twitterrank: finding topic-sensitive influential Twitterers. In: Proceedings of
the Third ACM International Conference on Web Search and Data Mining. ACM (2010)
40. Asharaf, S., Alessandro, Z.: Generating and visualizing topic hierarchies from microblogs:
an iterative Latent Dirichlet allocation approach. In: 2015 International Conference on
Advances in Computing, Communications and Informatics (ICACCI). IEEE (2015)
41. Quercia, D., Askham, H., Crowcroft, J.: TweetLDA: supervised topic classification and link
prediction in Twitter. In: The 4th Annual ACM Web Science Conference, pp. 247–250.
ACM, Evanston, Illinois (2012)
42. Onan, A., Korukoglu, S., Bulut, H.: LDA-based topic modelling in text sentiment
classification: an empirical analysis. Int. J. Comput. Linguistics Appl. 7(1), 101–119 (2016)
43. Cha, M., et al.: Measuring user influence in Twitter: the million follower fallacy. ICWSM
10, 10–17 (2010)
44. Silva, A., et al.: ProfileRank: finding relevant content and influential users based on
information diffusion. In: Proceedings of the 7th Workshop on Social Network Mining and
Analysis. ACM (2013)
45. Jiang, W., Wang, G., Wu, J.: Generating trusted graphs for trust evaluation in online social
networks. Future Gener. Comput. Syst. 31, 48–58 (2014)
46. Liu, B., Zhang, L.: A survey of opinion mining and sentiment analysis. In: Aggarwal, C.,
Zhai, C. (eds.) Mining Text Data, pp. 415–463. Springer, Cham (2012)
47. Balog, K., et al.: Expertise retrieval. Found. Trends Inf. Retr. 6(2–3), 127–256 (2012)
48. Yin, H.Z., et al.: Dynamic user modeling in social media systems. ACM Trans. Inf. Syst.
33(3), 10 (2015)
49. Abbasi, M.-A., Liu, H.: Measuring user credibility in social media, in social computing,
behavioral-cultural modeling and prediction. In: Greenberg, A., Kennedy, W., Bos, N. (eds.)
pp. 441–448. Springer, Berlin, Heidelberg (2013)
50. Yardi, S., et al.: Detecting spam in a Twitter network (2009)
51. Manyika, J., et al.: Big data: the next frontier for innovation, competition, and productivity.
McKinsey Global Institute (2011)
52. Chen, M., et al.: Open issues and outlook. In: Big Data, pp. 81–89. Springer International
Publishing (2014)
53. Lavbič, D., et al.: Traversal and relations discovery among business entities and people using
semantic web technologies and trust management. In: Databases and Information Systems
VII: Selected Papers from the Tenth International Baltic Conference, DB & IS 2012. IOS
Press (2013)
54. Herzig, J., Mass, Y., Roitman, H.: An author-reader influence model for detecting topic-
based influencers in social media. In: Proceedings of the 25th ACM Conference on
Hypertext and Social Media. ACM (2014)
55. Shen, W., Wang, J., Han, J.: Entity linking with a knowledge base: Issues, techniques and
solutions. IEEE Trans. Knowl. Data Eng. 27(2), 443–460 (2015)
56. Free Social Media Analytics Tools (2016). http://simplymeasured.com/free-social-media-
tools/
57. Li, C., et al.: Topic Modeling for short texts with auxiliary word embeddings. In:
Proceedings of the 39th International ACM SIGIR Conference on Research and
Development in Information Retrieval. ACM (2016)
Source Codes Classification Using a Modified
Instruction Count Pass
Omar Darwish1(&), Majdi Maabreh2, Ola Karajeh3,

and Belal Alsinglawi4
1
Computer Information System Department, Ferrum College,
Ferrum, VA, USA
odarwish@ferrum.edu
2
Computer Information System Department, Hashemite University,
Zarqa, Jordan
majdi@hu.edu.jo
3
Computer Science Department, Virginia Polytechnic Institute
and State University, Blacksburg, VA, USA
okarajeh@vt.edu
4
Western Sydney University, Parramatta, Sydney, NSW, Australia
b.alsinglawi@westernsydney.edu.au
Abstract. The vulnerability is a flaw in the system’s implementation which

may result in severe consequences. The existence of these flaws should be
detected and managed. There are several types of research which provide dif-
ferent solutions to detect these flaws through static analysis of the original
source codes. Static analysis process has many disadvantages, some of them are;
slower than compilation and produce high false positive rate. In this project, we
introduce a prediction technique using the output of one of the LLVM passes;
“InstCount”. A classifier was built based on the output of this pass on 500
source codes written in C and C++ languages with 88% of accuracy. A com-
parison between our classifier and Clang static analyzer showed that the clas-
sifier super performed to predict the existence of memory leak and Null pointers.
The experiment also showed that this classifier could be applied or integrated
with static analysis tools for more efficient results.
1 Introduction
The vulnerability is defined as a weakness or flaw in systems that make the systems
easy to be attacked or reduce its security level [1]. There are several types of vulner-
ability, some of them are related to memory such as; memory leak, null pointers and
buffer overflow. We concentrate on this project in the first two types. The problem of
memory leak is the allocation problem of some memory areas that didn’t release when
they no longer needed. For example, in OOP, such an object may be stored in a
memory and cannot be accessed by the running code. This memory area is allocated for
something unusable. A null pointer is defined as a pointer which points to the invalid
object. In other words, attempting to read from or write to a memory area which not
mapped or no longer used. This may lead to read rubbish values or null values. In both
cases, a crash may occur.
https://doi.org/10.1007/978-3-030-15035-8_88
898 O. Darwish et al.
Managing this problem can be done merely by manually auditing the source codes.
This approach may be valid if the code consists of a few lines or you have just a few
small codes. Moreover, this process needs good experience in programming and
software design. The security background is significant as well. This process is time-
consuming, and its results depend on the auditor and his/her skills. Due to the above,
other automatic techniques were introduced and now used to detect or predict the
existence of vulnerability. These techniques can be either static or dynamic. In the latter
one, an auditor will run the program and examine its behaviour through the runtime. In
[2], they provide a checking code technique with GCC compiler to detect those
potentially unsafe null pointers and memory leak. Static analysis is another technique
to detect the vulnerability.
In comparison with manual auditing, it can be efficient and save much time. In
general, the idea behind this technique is to depend on the predefined knowledge base
that contains keywords or functions which may cause the bug in the software. How-
ever, in practice, this technique has many limitations and disadvantages. The high false
positive is one of the weaknesses in tools which used static analysis. The generated
reports contain many unsafe programs, but in reality, this is not true. A flawfinder and
LLVM clang static analyzers are the example of those tools [3, 4].
2 LLVM Environment
LLVM is a compiler infrastructure that provides a middle layer of compiling process. It

takes the Intermediate Representation (IR) from the compiler and generates the opti-
mized image of this IR. The optimized IR can be transformed into machine code [5].
LLVM provides different passes or optimizers. Some of them are just for analysis, and
the other may transform the source code into another optimized form. One of them is
“instructions count” or “InstCount” pass. This pass returns the total number of
instructions which belong to the following types; LOAD, STORE, ALLOCATE,
CALL, INVOKE and more. These instructions play a significant and essential role in
code execution; memory and time.
3 Instruction Count Pass
This pass is one of many passes developed by LLVM team. It is an analysis pass and
doesn’t change the source code but gives some statistics about the code. These values
are so important and give the programmer an essential guide about his/her code exe-
cution. In this section, the output of this pass will be discussed in details.
3.1 Terminator Instructions

This type contains the following instructions: ‘ret’, this instruction is frequently used in
programs where it returns the control flow from the function back to the caller. If the
flow back to the caller, then the execution will be continued after the call statement and
a value (if any) shall set the caller value. ‘br’ instruction, transfers the execution into
Source Codes Classification Using a Modified Instruction Count Pass 899
different basic block based on condition(s). This happens when, for example, a con-
ditional statement used in the code.
In the same fashion, a ‘switch’ instruction can be used. The difference between ‘br’
and ‘switch’ is that switch can use many ‘br’ as it transfers the execution from one
place to another. Similarly, the ‘invoke’ instruction causes control to transfer to the
specified function, with the possibility of control flow transfer to either the
‘normal’ label or to ‘exception’ label. The normal label will be used if the caller uses
the ‘ret’ instruction. Without a successor, ‘resume’ is termination instruction.
If a source code contains a statement that will not be executed anyway, this means
these statements are unreachable. ‘Unreachable’ instructions can be counted via this
pass. This is a good indication that these statements should be removed or managed in
such a way.
3.2 Binary Operations

This type of instruction needs two operands, and they do most of the computation
operations in any program. They produce a single value as a result. This type consists
of: ‘Add’, which return the sum of its two operands. This instruction is appropriate for
both signed and unsigned integers in LLVM. While ‘add’ is just for integers values,
‘fadd’ is an instruction used to sum two floating point numbers into float result. For
subtraction, ‘sub’ produces the difference of its operands as an integer value. The ‘fsub’
is used with float numbers. The ‘mul’ and ‘fmul’ are used for multiplication for integer
and float numbers, respectively. Regarding division instructions, there are three
instruction types which deal with signed ‘sdiv’ and unsigned integers ‘udiv’, and one
used with float numbers, ‘fdiv’.
The ‘srem’ and ‘urem’ are used for the remainder from a division operation. The
former is for signed, and the latter is for unsigned integers. For float numbers, the
reminder can be computed by using ‘frem’ instruction.
3.3 Bitwise Binary Operation

These instructions are used in bit-twiddling operations. They take two operands and
return a single value. The data type of the results depends on the data types of the
operands. This type of instruction consists of the following: ‘Shl’, which return the first
operand as shifted to the left n bits, where n is the second operand. In the same way,
‘lshr’ returns the first operand shifted to the right n bits where n is the second operand.
Both of the above instructions fill the shifted space by zeros. The ‘ashr’ instruction
returns the first operand shifted to the right n bits with signed extension. It always
applies the arithmetic shift right operation. The rest of the instructions of this type are
those well known in computer science; ‘and’, ‘or’, and ‘xor’.
3.4 Vector Operations

The ‘extractelement’ is used to extract a specific value from a vector in a specific index.
It takes two operands; the first is the vector and the second is the position.
‘shufflevector’ constructs a permutation of two vectors.
3.5 Aggregate Operations

The aggregate operations are similar to a vector, but on aggregate values, the
‘extractelement’ gets a specific value from aggregate values. The first operand will be
an array or struct, and the second one is for the position. Conversely, ‘insertelement’ is
used to insert the value into aggregate values.
3.6 Memory Access

This section describes how LLVM deal with memory by specifying the instructions.
The ‘alloc’ allocates memory on the stack frame of the function currently executed. The
allocated area will be automatically released after the execution. ‘load’ instruction is
used to fetch values from memory. Conversely, ‘store’ instruction is used to write
values to memory. There are many instructions related to memory addressing, but here
we focused just on the above instructions.
3.7 Other Operations

In this section, several instructions will be described as those are not classified into the
above. These instructions are so important and play a major role in our work. ‘icmp’
returns a Boolean value of a comparison between two integers, vectors or pointers. For
float numbers or vectors, the ‘fcmp’ is used instead of ‘icmp’. The ‘phi’ is used to
represent the PHI nodes in SSA graph. ‘select’ is used to choose one value based on a
condition. To call functions, a ‘call’ instruction is used and flow will be transferred into
a different place. Several instructions can be collected and computed by this pass, but in
this research, we are focusing on the previously mentioned instructions.
4 Related Work
This research is considered a multidisciplinary topic. It contains security and data

mining. Here, we will focus on some works in literature related to vulnerability pre-
diction and detection in source codes. The static analysis of source codes has several
approaches. The simplest one is the pattern matching, where the tools just examine the
occurrence of such pattern using the keyword of functions names such as strcpy().
A flawfinder tool is a well-known one in the literature that has a large built-in database
contains vulnerable patterns [3] to check against the C source codes. Similarly and for
other languages, ITS4 [6] can do another check beside the analysis to make sure that a
suspected statement has a vulnerability or not. Another approach is the lexical analysis,
where the source code transformed into a series of tokens without white spaces and
then checked by one of the above tools.
In the prediction of vulnerable software, and to find the relation between the
software complexity and the bugs in Mozilla javascript engine, Shin and Williams [7]
find a weak correlation between them, and the results cannot be generalized.
Chowdhury and Zulkernine [8] examine if the coupling and cohesion metrics can be
used to predict the existence of vulnerability. Their experiment over 52 version of
Mozilla Firefox showed that it is possible in around 72% of the time. This is based on
the classifier built using decision tree. Hata, Mizuno et al. [9], provide an approach
using text analysis to predict flaws in software. They applied the approach on five open
source eclipse projects and got just 40% of precision and 80% of recall. Using text
mining, Aversano et al. [10] built a predictor to check the newly added code in such
software contains bug or not. The KNN technique was used and applied over two open
java source. The built tool got 59% precision and 69% of recall. The above is a sample
of the related work in bug prediction in software. All of them use the high-level source
code analysis to build a tool for vulnerability detection. [11] performed an early study
to predict vulnerability using text mining on the source code. Our work has the same
target, but we try to do the same thing using low-level codes.
5 Methodology
For the research purposes, 1000 source codes written in C and C++ language are
collected from different sources. We use the benchmark of NIST [12], and around 700
codes from math library and planet source archives websites [13]. These codes are
compiled using Clang built in Ubuntu 14.0. There are many source codes contains
syntax errors. We try to fix them, but it is time-consuming and needs too much effort as
some of them include more than one error. In our dataset, we use 250 benchmark codes
from NIST and other 250 codes from the other sources. To make sure that the latter 250
are free of bugs, we have used the flawfinder tool, and we found that most of them are
clean. The buggy codes are replaced by others where the flawfinder decide those as
clean codes. The second step in this research is to generate the bitcodes of the dataset.
We use LLVM commands for this issue. After that, the “instcount” pass was modified
to get a specific output format, then applied over the dataset. Every source code will get
another txt file contains its analysis results. An around 250 LOC matLab analyzer was
built to collect the files and parse the content and unify them in one csv file to be ready
for data mining WEKA tool. Figure 1 below shows the processes of this research.
Fig. 1. Research phases

6 Results and Discussion
The instcount model was trained on 66% of the 500 records and tested on the other
ones. These numbers are default values in WEKA. A decision tree generated from
WEKA tool was implemented. It contains 12 attributes from originally 40 attributes
that WEKA found them as good features to classify the 500 source codes into three
classes; clean, the vulnerability in the pointer, and vulnerability of memory leak. It is
clear that the classification process depends mainly on the values of basic blocks and
Phi. Also, the unreachable instructions have a major role in building the decision tree.
The confusion matrix in Fig. 2 (Appendix) shows that the classifier accuracy is around
88%. This means that this classifier can be applied in real applications. The classifi-
cation process based on the low level gives a strong indication that we can predict the
existence of vulnerability based on the relation between the instructions counts and
values. In comparison with static analysis tools, we found that our classifier super
performed the LLVM Clang static analyzer. We use all the benchmark codes from
NIST to assess our classifier, flawfinder, and clang analyzer. Table 1 shows that LLVM
clang analyzer needs more effort to improve its accuracy and efficiency in vulnerability
detection. We also apply the WEKA decision tree technique to classify the vulnerable
software itself into two classes we are working on; null pointer and memory. The
classification process showed that we could build an efficient classifier to differentiate
between vulnerability types. Figure 3 (Appendix) indicates that the constructed clas-
sifier gets around 94% of accuracy. After building the decision tree and generate the
rules, the InstCount pass in LLVM was modified, and the rules were added. The output
of the new version contains the decision of the classifier about the vulnerability exists
in the input source code. Figure 4-a, b, c (Appendix) shows the output of sample files
of source codes.
Table 1. Comparison with other tools

Analyzer/Passer Correct Not correct
Clang analyzer LLVM 235 92
Flawfinder 321 6
New InstCount Passer 281 46
7 Obstacles and Difficulties
In this research, we spent too much time in dataset collecting. Many difficulties are
encountered because many of the downloaded files have a syntax error(s). Some file
can be compiled in the windows environment, but not in Linux, as they contain some
pure windows header files and they need the effort to convert them into Linux like
environment. In order to get reliable results, the dataset should be as much as possible.
The collecting process is not easy and needs effort and time.
In this project, we exploit the results of LLVM pass to predict the existence of vul-
nerable software. We depend on the low-level instructions. We use WEKA tool to
build a decision tree and generate the rules. Our classifier has 88% of accuracy. Based
on the results, the classifier super performed the LLVM Clang Static Analyzer in
prediction the buggy source codes. However, both of them are using a different
technique; we argue that our classifier can be considered as enhancement of the Clang
analyzer and we can use both of them together for more efficient results.
For the future plan, we will study these results in depth to find the relation among
the instruction numbers that gives the pattern of buggy codes. Also, we are going to
enrich the dataset for both vulnerable and clean source codes. Moreover, we will use
advanced data mining techniques to find more strong rules based on the instructions.
Appendix
Fig. 2. Confusion matrix.
Fig. 3. Confusion matrix for vulnerable source codes.

(a)
(b)
(c)
Fig. 4. Sample of the output of new instcount pass.
References
1. Hughes, J., Cybenko, G.: Three tenets for secure cyber-physical system design and
assessment. In: Cyber Sensing 2014, vol. 9097, p. 90970A. International Society for Optics
and Photonics (2014)
2. Eigler, F.C.: Mudflap: Pointer use checking for C/C++ (2003)
3. Flawfinder. http://www.dwheeler.com/flawfinder/
4. Clang Static Analyzer. http://clang-analyzer.llvm.org/
5. The LLVM Compiler. Infrastructure. http://llvm.org/
6. Viega, J., Bloch, J.T., Kohno, Y., McGraw, G.: Its4: a static vulnerability scanner for C and
C++ code. In: 16th Annual Conference on Computer Security Applications, ACSAC 2000,
pp. 257–267. IEEE (2000)
7. Shin, Y., Williams, L.: Is complexity really the enemy of software security? In: Proceedings
of the 4th ACM Workshop on Quality of Protection, pp. 47–50. ACM (2008)
8. Chowdhury, I., Zulkernine, M.: Using complexity, coupling, and cohesion metrics as early
indicators of vulnerabilities. J. Syst. Architect. 57(3), 294–313 (2011)
9. Mizuno, O., Ikami, S., Nakaichi, S., Kikuno, T.: Spam filter based approach for finding fault-
prone software modules. In: Proceedings of the Fourth International Workshop on Mining
Software Repositories, p. 4. IEEE Computer Society (2007)
10. Aversano, L., Cerulo, L., Del Grosso, C.: Learning from bug-introducing changes to prevent
fault prone code. In: Ninth International Workshop on Principles of Software Evolution: in
Conjunction with the 6th ESEC/FSE Joint Meeting, pp. 19–26. ACM (2007)
11. Scandariato, R., Walden, J., Hovsepyan, A., Joosen, W.: Predicting vulnerable software
components via text mining. IEEE Trans. Softw. Eng. 40(10), 993–1006 (2014)
12. National Institute of Standards and Technology. http://www.nist.gov
13. Mathematics Source Library C & ASM. http://www.mymathlib.com
Predictive Analytics and Deep Learning
Techniques in Electronic Medical Records:
Recent Advancements and Future Direction
Belal Alsinglawi(&) and Omar Mubin
Western Sydney University, Parramatta, Sydney, NSW, Australia

{B.Alsinglawi,O.Mubin}@westernsydney.edu.au
Abstract. The demands on medical services are increasing rapidly in the global
context. Therefore, handling beds availability, identifying and managing the
length of stay (LOS) is creating persistent needs for the physicians, nurses,
clinicians, hospital management, and caregivers in the public hospital admis-
sions and the private hospital admissions. Health analytics provides unprece-
dented ways to predict trends, patients’ future outcomes, knowledge discovery,
and improving the decision making in the clinical settings. This paper reviews
the state-of-the-art machine learning, deep learning techniques and the related
work in relation to the length of stay common hospital admissions. Research
trends and future direction for the forecasting LOS in medical admissions are
discussed in this paper.
1 Introduction
The advancement in health informatics systems has shown an unprecedented amount of

rich clinical data of the patients’. Electronic health records (EHRs) or the “Electronic
Medical Records (EMR)” have generated feasible opportunities to help clinicians to
understand and address broad range of questions in medicine [1]. The widespread of
the EHRs adoption in the medical informatics and decision support systems have
opened the doors to researchers, medical professional and healthcare providers to
improve the healthcare domain, support clinical decision making, and provide better
clinical outcomes. Typically, EMR data contains the patient’s medical history, diag-
noses, medications, treatment plans, immunization dates, allergies, radiology images,
laboratory, test results [2], and all administrative medical data such as LOS. Within
electronics health records, the length of stay (LOS) is a significant quality metric that
aims to determine the time of how long the patients spend in the hospital from the time
they are admitted to into hospital to the time of discharge.
The LOS is measured in months, weeks, days, and hours. LOS metric measure is
essential for hospitals and medical centers to decide the type of diagnosis for each
group within the particular medical department (e.g. ICU). The high demand on the
EHRs in the medical field has encouraged many researchers to build predictive health
analytics modelling from electronic health records (e.g. regression, classification, and
clustering methods). The LOS Predictive modellings play a significant role in facili-
tating hospitals operational success during hospital various admission and emergency

https://doi.org/10.1007/978-3-030-15035-8_89
908 B. Alsinglawi and O. Mubin
departments (e.g. Acute myocardial infarction in the ED). Machine learning techniques
are expediting the processes for the artificial intelligence (AI) in the clinical field and
helping the researchers in this domain to construct predictive medical modelling and
health analytical tools that can support clinical decision making.
Machine learning and recently Deep Learning models can work effectively with
EHRs in an automatic and collective process [3]. As a result, this will help the
researchers and the data scientists to collect massive records of data, including patients’
health records, behavioral data, temporal data and time series data; therefore, better
intelligent diagnostic predictions from the data can be achieved. Predicting the number
of days which patients are required to stay at hospitals can be beneficial for hospitals,
patients and caregivers. Consequently, this will lead to preventing the late discharge or
early discharge for particular medical problems that require a prolonged LOS. In
prediction learning modellings, whether supervised learning or unsupervised models
can best predict the particular medical case, or even hybrid approach it is necessary to
deal with the complex medical case that requires appropriate data representation prior
to the prediction process.
The motivation of this work relies on the fact that highlighting and discussing the
most recent advancement in predictive analytics algorithms may lead to a better
understanding of current techniques of the AI in LOS EHR predictive tasks. This
comprises studying these techniques in the literature review, research challenges and
the future research direction for forecasting LOS predictions problems. Therefore, this
paper aims to summarize the state-of-the-art machine learning and deep learning
techniques related works for predicting LOS in common hospital admissions. These
works are group into five main categories. Acute Hospital Admission, Intensive Care
Unit (ICU), Emergency Department ED, Pediatric Emergency Department (PED),
General Hospital Admission, and Cardiac Surgery.
2 Predictive Modeling in EHR Length of Stay

(LOS) in Hospital Admissions
Many research projects [4, 5] have endeavored to address LOS prediction at the
emergency department, ICU and the different types of hospital admission considering
LOS as a prediction problem using regression analysis models, while other research
works [6–8] have taken into account the classification models in the LOS prediction
problems, meanwhile they are fewer types of research exploiting clustering-based
models in LOS. These models were examined in various clinical settings for fore-
casting of length stay tasks and to find the best predictive technique in Acute Medicine,
Intensive Care Unit (ICU), Emergency Department ED, Pediatric Emergency Depart-
ment (PED), General Hospital Admission, and Cardiac Surgery. These are the common
predictive tasks in forecasting LOS in hospitals admission. Each predictive task has
different affecting factors which may influence the average length of stay each one
(short LOS, medium LOS, and long LOS or “prolonged LOS”).
Predictive Analytics and Deep Learning Techniques 909
2.1 Regression Analysis Related Work

Regression analysis based methods are popular due to their characteristics to predict the
value of the target such as numerical values. From regression analysis based methods,
we can build predictive models from one or more predictors (numerical values or
categorical variables).
Numerous works [4, 5, 9–11] in the literature review implemented Regression
analysis to forecast length of stay in various hospital admissions. For instance, Logistic
regression (LR), Multivariate binary Logistic Regression, Poisson Regression model
based techniques are used in the emergency department prediction tasks and PED
admission. In the Intensive Care Unit (ICU) prediction tasks, some research works
studied regression analysis models such as [12, 13]. In Acute hospital admission,
regression analysis was implemented as the most common predictions model [14–16].
Table 1, shows regression analysis based works in the main hospital admission pre-
dictive tasks.
Table 1. Represents common regression analysis hospital LOS related works.

Model Example Benefits of regression Drawbacks of the regression
analysis-based models analysis-based models
Regression [4, 5, 9–16] – Easy for probabilistic – Unable to perform well
analysis models interpretation when there are multiple and
– Fast to train nonlinear classification
– Avoid overfitting problems
– Easy to update with – Not efficient to capture
stochastic gradient descent complex relationships
between variables
2.2 Classification Supervised Learning Related Work

In the classification supervised models, there are classes (targets/labels), or categories
which are used in the process of predicting class(s) of given data points. The supervised
classification predictive models are performed by the task of approximating a mapping
function (f) from input variables (X) to discrete output variables (y).
Support Vector Machine (SVM) is used in prediction hospital admission problems
such LOS hospital admission such as (PED) [17], and ICU after (Cardiac and Heart
Surgery) [6]. The Random Forest (RF) which is one of the well-known supervised
classification algorithm and it is commonly used in Predicting LOS in PED [17],
General Hospital Admission [8], and the LOS After Surgery (Common Medical Sur-
gery Admission) [18]. The k-nearest neighbour (kNN) and Naïve Bayes related works
are applied in the LOS (After Surgery “Common Medical Surgery Admission”) such as
[19, 20]. Table 2, illustrates supervised learning models, where it compares the benefits
and drawback of these modes in Hospital admission LOS.
Table 2. Represents a comparison between the common supervised learning models in LOS
Hospital Admissions common predictions problems.
Classification Example Model advantages Model disadvantages
model
Support [6, 17] – Fast to predict outcomes – Not efficient in large
Vector – Efficient in non-linear dataset (require a lot of time
Machine separation problems for training data)
(SVM) – Computation speed and – Not efficient when target
memory classes are overlapping
K-Nearest [19] – Simple algorithm and – Features selection problem
Neighbors easy to implement – Takes a lot of memory and
(kNN) – Learning is simple CPU to calculate nearest
(almost no learning at all) neighbors in a large dataset
Naive Bayes [20] – Fast to predict – Not suitable if categorical
– Useful in multi-class variable has Zero probability
prediction – Not efficient if predators are
– Efficient in categorical not completely independent
input variables
Random [8, 17, 18] – Useful in non-linear – Computational cost
Forest prediction problems – Takes time to build and
– Useful in feature create
engineering (can find
important features in trained
dataset)
2.3 Unsupervised Learning Related Work

Unsupervised learning is used when we have only input data (X), and there is no
corresponding output (variables). Unsupervised learning aims to model the underlying
structure or the distribution in the data to learn more about the data. Unsupervised
learning is unlike supervised learning where there are no correct answers, and also there
is no teacher. Unsupervised algorithms are left to their devises to discover and present
the interesting structure in the data. In unsupervised learning, the problems can be
further grouped into clustering and association problems.
Clustering Models, K-Means Clustering is used when there is unlabeled data that is
not defined in categories or groups. The goal of cluster analysis in K-Means (Eq. 2)
algorithm is to cluster or group the observations to a subset according to their responses
on the multiple variables. For instance, few research implemented K-means such in
Intensive Care Unit (ICU) length of stay.
2.4 Deep Learning Models in LOS Hospital Admissions

Deep Neural Networks or (Deep Learning) algorithms have tremendous benefits in
feature representation from raw data. The main benefit of using deep learning-based
techniques is that their ability to learn feature representation automatically from the raw
EMR data and this will help to reduce manual efforts which are made in the case of the
traditional machine learning techniques in clinical settings. Further, deep learning
techniques can engineer features which are suitable for many clinical and medical tasks
or customize specific features for specific tasks [21]. In general, deep learning based
models require large data to work efficiently. In other words, they need enough data to
train deep learning models in various prediction problems.
Recurrent Neural Network (RNN) is a common deep learning technique. RNN
(Fig. 1) is a type of neural networks which is an appropriate choice when clinical data
is sequentially ordered (time-series or natural language), especially that RNNs are
designed to deal with such long-range temporal dependencies [22]. Recently,
researcher started to exploit the immense benefits of deep learning based techniques in
EMR context. For example, Che et al. [23] applied RNN in time-series classification on
multiple clinical prediction tasks (including the length of stay) extracted from three
openly available databases (UCI, MIMIC III, and PhysioNet). Similarly, Harutyunyan
et al. [24] applied RNN using public database (MIMIC III) to predict the length of stay
in the problem of the multiple time series event in clinical settings.
Fig. 1. Recurrent neural network architecture.
The LSTM neural network is a type of the RNN which can learn order dependences
in sequential prediction problems. The LSTM consists of three gates (Eqs. 1, 2, and 3);
the first gate is the forget state (ft), the second gate is the input gate (It) and the output
gate (Ot), where each gate has different sets of weights. The following equations
represent the three gates and three states:
ft = ϭ(Wf St-1 + Wf Xt) ð1Þ
It = ϭ(Wt St-1 + Wi Xt) ð2Þ
Ot = ϭ(Wo St-1 + Wo Xt) ð3Þ
Numerous research studied the LSTM performance in the clinical domain. For
instance, Fang et al. [1] compared the performance of class models (Logistic Regres-
sion (LR) and Feed-Forward (MLP)) with LSTM, and Multitask LSTM models in LOS
prediction task. The results indicated that LSTM and Multitask LSTM outperformed
traditional learning (shallow learning). Comparably, Rajkomar el at [25] applied deep
learning LSTM to predict the length of stay of hospitalized patients in the first 24 h of
admission.
2.5 Natural Language Processing (NLP)

The natural language processing (NLP) is the process of deriving meaningful infor-
mation from natural language text (e.g. clinical notes, doctors discharge summaries
etc.). The two components of NLP are the natural language understanding and the
natural language generation. The natural language understanding refers to mapping the
given input into natural language into a useful representation and analyzing those
aspects of language. The natural language generation is the process of producing
meaningful phrases and sentences in the form of natural language some internal rep-
resentation. There are two techniques to understand natural language; the first is
Syntactic Analysis (Syntax), and the second is Semantic Analysis (Semantic). The
syntactic analysis involves analyzing natural language conforming into rules of formal
grammar. The syntactic analysis is responsible for assigning semantic structure to text.
The Semantic Analysis refers to the process of understanding the meaning of the
interpretation of words, sentence structure, and signs. Recently, the researcher started to
take advantage of deep learning techniques (LSTM, RNN etc.) in NLP algorithms
using the deep generative models while solving the complex predictive NLP tasks.
Lately, Natural Language Processing (NLP) models provided new research insight
in the domain of Hospital LOS in medical cases admission [25]. Taking advantage of
the NLP models will aid to construct LOS predictive models from the rich clinical
notes, and ultimately will help to facilitate the mission to reduce waiting time during
hospital admissions.
3 Future Direction
With the recent advancement in deep learning models in the medical domain, however,
deep learning models are still in the early development stage in EMR context with the
focus on the length of stay in the common hospital admission prediction tasks. Since
there are little few attempts to take advantage of the deep learning techniques in this
field, the future directions and trends can be pointed to the following research interest:
– Maximize the benefits of the current models such as DNN, RNN, LSTM and
implements Unsupervised (e.g. Deep Belief Network, Autoencoder, etc.) and Rein-
forcement deep learning models to ease and build a more efficient model to achieve
promising results in various hospital admission prediction tasks including hospital
length of stay, discharge, the risk of re-admission, and future patients’ outcomes.
– Deep learning based techniques have tremendous benefits for Big Data analytics
such the use of deep neural network layers of hidden variables, and the automatic
features learning (pre-processing, feature, extraction and feature selection.). In the
domain of deep learning and big data health analytics, there are common techniques
which were recently adopted in hospital admissions and prediction tasks such as the
LSTM and RNN. Recently, researchers started to exploit the benefits of deep
learning unsupervised techniques in big data analytics such as (Auto-encoders,
Boltzmann Machines (RBMs), and Deep Belief Networks (DBN)). Therefore, the
adoption of big data analytics remains unexplored research area in the domain of
EHR for the hospital length of stay prediction problems.
– During hospital inpatient, Informative analytics in IoT data is necessary to make

useful better decisions and optimize hospital operations. Applying real-time ana-
lytics to monitor medical conditions in hospital admission and respond accordingly
is deemed a significant task in the clinical setting with the use of IoT analytics [26].
4 Conclusion
Previous research tells us identifying the right predictive learning models that can suit
specific medical cases in forecasting length of stay for is rendered as a significant task
in hospital admissions. Certainly, this is considered as a research challenge especially
in deep learning models and big data analytics which require further investigations to
provide promising outcomes to forecast hospital length of stay for the target medical
prediction task.
References
1. Tang, F., Xiao, C., Wang, F., Zhou, J.: Predictive modeling in urgent care: a comparative
study of machine learning approaches. JAMIA Open (2018)
2. What is an electronic health record (HER) (2018). https://www.healthit.gov/faq/what-
electronic-health-record-ehr
3. Gunasekar, S., Ho, J.C., Ghosh, J., Kreml, S., Kho, A.N., Denny, J.C., Malin, B.A., Sun, J.:
Phenotyping using structured collective matrix factorization of multisource ehr data. arXiv
preprint arXiv:1609.04466 (2016)
4. Liu, V., Kipnis, P., Gould, M.K., Escobar, G.J.: Length of stay predictions: improvements
through the use of automated laboratory and comorbidity variables. Medical care, pp. 739–
744 (2010)
5. Hosseininejad, S.M., Aminiahidashti, H., Pashaei, S.M., Khatir, I.G., Montazer, S.H.,
Bozorgi, F., Mahmoudi, F.: Determinants of prolonged length of stay in the emergency
department; a cross-sectional study. Emergency 5(1) (2017)
6. Hachesu, P.R., Ahmadi, M., Alizadeh, S., Sadoughi, F.: Use of data mining techniques to
determine and predict length of stay of cardiac patients. Healthc. Inform. Res. 19(2),
121–129 (2013)
7. Maharlou, H., Niakan Kalhori, S.R., Shahbazi, S., Ravangard, R.: Predicting length of stay
in intensive care units after cardiac surgery: comparison of artificial neural networks and
adaptive neuro-fuzzy system. Healthc. Inform. Res. 24(2), 109–117 (2018)
8. Elbattah, M., Molloy, O.: Using machine learning to predict length of stay and discharge
destination for hip-fracture patients. In: Proceedings of SAI Intelligent Systems Conference,
pp. 207–217. Springer (2016)
9. Azari, A., Janeja, V.P., Levin, S.: Imbalanced learning to predict long stay emergency
department patients. In: 2015 IEEE International Conference on Bioinformatics and
Biomedicine (BIBM), pp. 807–814. IEEE (2015)
10. Vermeulen, M.J., Guttmann, A., Stukel, T.A., Kachra, A., Sivilotti, M.L., Rowe, B.H.,
Dreyer, J., Bell, R., Schull, M.: Are reductions in emergency department length of stay
associated with improvements in quality of care? a difference-in-differences analysis. BMJ
Qual. Saf. 25(7), 489–498 (2016)
11. Barak-Corren, Y., Fine, A.M., Reis, B.Y.: Early prediction model of patient hospitalization
from the pediatric emergency department. Pediatrics 139(5), e20162785 (2017)
12. Carter, E.M., Potts, H.W.: Predicting length of stay from an electronic patient record system:
a primary total knee replacement example. BMC Med. Inform. Decis. Making 14(1), 26
(2014)
13. Zaccardi, F., Webb, D.R., Davies, M.J., Dhalwani, N.N., Gray, L.J., Chatterjee, S., Housley,
G., Shaw, D., Hatton, J.W., Khunti, K.: Predicting hospital stay, mortality and readmission
in people admitted for hypoglycaemia: prognostic models derivation and validation.
Diabetologia 60(6), 1007–1015 (2017)
14. Juma, S., Taabazuing, M.M., Montero-Odasso, M.: Clinical frailty scale in an acute
medicine unit: a simple tool that predicts length of stay. Can. Geriatr. J. 19(2), 34 (2016)
15. Magalhães, T., Lopes, S., Gomes, J., Seixo, F.: The predictive factors on extended hospital
length of stay in patients with ami: laboratory and administrative data. J. Med. Syst. 40(1), 2
(2016)
16. Cereda, E., Klersy, C., Pedrolli, C., Cameletti, B., Bonardi, C., Quarleri, L., Cappello, S.,
Bonoldi, A., Bonadeo, E., Caccialanza, R.: The geriatric nutritional risk index predicts
hospital length of stay and in-hospital weight loss in elderly patients. Clin. Nutr. 34(1),
74–78 (2015)
17. Combes, C., Kadri, F., Chaabane, S.: Predicting hospital length of stay using regression
models: application to emergency department. In: 10ème Conférence Francophone de
Modélisation, Optimisation et Simulation-MOSIM’14 (2014)
18. Chuang, M.T., Hu, Y.H., Lo, C.L.: Predicting the prolonged length of stay of general
surgery patients: a supervised learning approach. Int. Trans. Oper. Res. 25(1), 75–90 (2018)
19. Salah, H.: Predicting Inpatient Length of Stay in Western New York Health Service Area
Using Machine Learning Algorithms. Ph.D. thesis, State University of New York at
Binghamton (2017)
20. Muhlestein, W.E., Akagi, D.S., Davies, J.M., Chambless, L.B.: Predicting inpatient length of
stay after brain tumor surgery: developing machine learning ensembles to improve predictive
performance. Neurosurgery (2018)
21. Ching, T., Himmelstein, D.S., Beaulieu-Jones, B.K., Kalinin, A.A., Do, B.T., Way, G.P.,
Ferrero, E., Agapow, P.M., Zietz, M., Hoffman, M.M., et al.: Opportunities and obstacles for
deep learning in biology and medicine. J. R. Soc. Interface 15(141) (2018)
22. Shickel, B., Tighe, P., Bihorac, A., Rashidi, P.: Deep EHR: a survey of recent advances in
deep learning techniques for electronic health record (EHR) analysis. arXiv preprint arXiv:
1706.03446 (2017)
23. He, Z., Purushotham, S., Cho, K., Sontag, D., Liu, Y.: Recurrent neural networks for
multivariate time series with missing values. Sci. Rep. 8(1), 6085 (2018)
24. Harutyunyan, H., Khachatrian, H., Kale, D.C., Galstyan, A.: Multitask learning and
benchmarking with clinical time series data. arXiv preprint arXiv:1703.07771 (2017)
25. Rajkomar, A., Oren, E., Chen, K., Dai, A.M., Hajaj, N., Hardt, M., Liu, P.J., Liu, X.,
Marcus, J., Sun, M., et al.: Scalable and accurate deep learning with electronic health
records. NPJ Digital Med. 1(1), 18 (2018)
26. Dimitrov, D.V.: Medical internet of things and big data in healthcare. Healthc. Inform. Res.
22(3), 156–163 (2016)
Big Data Analytics for Electricity Price
Forecast
Ashkan Yousefi1, Omid Ameri Sianaki1(&), and Tony Jan2

1
Victoria University Sydney, Sydney, Australia
{ashkan.yousefi,omid.amerisianaki}@vu.edu.au
2
Melbourne Institute of Technology, Melbourne, Australia
tjan@mit.edu.au
Abstract. Electricity Price forecast is a major task in smart grid operation.

There is a massive amount of data flowing in the power system including the
data collection by control systems, sensors, etc. In addition, there are many data
points which are not captured and processed by the energy market operators and
electricity network operators including gross domestic product, the price of fuel,
government policy and incentives for renewable and green energy sectors as
well as impacts on new technologies such as battery technology advancement
and electric vehicles. In this study, data points from 2001 to 2017 were collected
and 78 data points are considered for analyses to select the highly-correlated
features which could potentially affect the electricity price. In the first step, a
comprehensive correlation method using Pearson Correlation Coefficient is
applied to find the points which are correlated with the electricity price. In the
next step, the correlated data is fed to the machine learning algorithm for price
forecast. The algorithm results were tested in the historical data in California and
the outcomes were satisfactory for the three years forecast. The combination of
featured selection and machine learning is giving superior outcomes than the
traditional methods.
1 Introduction
One of the main applications of smart grid is the ability to provide the flexible load
curve and shift the load from the peak time. To be able to achieve this goal, dynamic
pricing would be a great help. Accurate price forecast can help the customers to make a
decision for their electricity consumption and adjust their electricity consumption
pattern based on accurate forecast. This quantitative and accurate decision making not
only helps the customers to optimize their electricity consumptions but also helps the
electricity market participants and the operators to run the market more efficiently. The
features which are going to influence the network are very diverse and having an
algorithm which can extract the important features are crucial to determine the price of
the electricity. The good news is that large installations of the smart meters provide
very comprehensive historical data which can be used for electricity price forecast in
short and long term [1–3].
During the current research the electricity price forecast researches was reviewed
and the various applied methods are analyzed. Summary of the sum of the analyzed

https://doi.org/10.1007/978-3-030-15035-8_90
916 A. Yousefi et al.
researches presented in the following paragraph. Many of the publications related to the
electricity price forecast applied time series analysis and the machine learning methods
for electricity price forecast. A hybrid model for predicting the electricity price is
implemented by Varshney to predict day ahead electricity market based on the
extracted featured such as temperature and demand information [4]. A probabilistic
methodology combined with the wavelet neural networks is utilized to forecast the
hourly price with considering the uncertainty [5]. A markov chain model was used to
calculate the electricity pricing based on welfare with considering a tradeoff between
the customers and the utility companies [6]. The hourly electricity price forecast is
improved by applying multilayer neural network. The applied neural network in [7]
achieved the promising result and improved the accuracy of the forecast.
Time series analysis also used extensively for the purpose of forecasting the
electricity price. One of the most famous methods used for times series analysis is
called auto-regressive integrated moving average (ARIMA) [8]. Feature engineering is
one of the key requirements to determine the electricity price as the accurate forecasting
the electricity price demands feeding the relevant historical data with most correlated
features to the electricity price. The research referenced in [9–11] applied multi-
variable mutual information, interaction gain as well as C4.5 algorithms for feature
extractions. Traditional studies applied the decision trees as well as artificial neural
networks and according to [12] and [13] the traditional methods are not able to leverage
the big data values for the price forecast very efficiently.
2 Analysis and Methodology

2.1 Data Collection and Cleaning
Data from the US Energy Information Administration’s OpenData database [14] (one
of the earliest completely open government data portals) was utilized for this project
simulations and study. Querying for the data in CSV format through their API, minor
cleaning was conducted (column names, local average for missing values when rea-
sonable), and datatables were merged to get consistent data from 78 different data
points for every month between January 2001 and August 2017.
2.2 Feature Correlation Analysis

Correlation analysis was conducted via generation of a Pearson correlation matrix on
the cleaned data using the Pandas Python package [15] in order to find the data that was
highly correlated with the retail price of electricity. We summarize highly correlated
features, and present interesting insights. The Fig. 1 illustrates some of the highly
correlated features from the massive data set we have processed with the price of
electricity.
Big Data Analytics for Electricity Price Forecast 917
2.3 Time Series Prediction

Moving Average [16] model was trained via the autoregression model in the Stats
models Python Package to produce the three years forecast for the electricity price.
Following the promising performance of the S-ARIMA model in capturing seasonality
of energy data, the use of a simple machine learning model on the ARIMA-projected
values of features in order to predict electricity price was considered. Several models
were trained on the hand-picked features with the goal of achieving a lower mean
squared error than the 5-year time series predictor. The methodology proceeded based
on following steps:
1. Used Time Series analysis to make predictions for values of highest-correlated
features in the future.
2. Trained several models with true values of highest-correlated features in the present.
3. Applied coefficients of tuned algorithm as weights to projected values of highest-
correlated features to obtain a projected future retail cost of electricity.
3 Numerical Study
3.1 Data Collection and Cleaning

Despite the ample availability of public energy data on the web, obtaining consistent
and reliable energy price data for a specific geographic region was a major challenge.
Data was gathered from tens of sources and merged, then compared with that published
by each hub belonging to energy companies in California. Inconsistencies were found
in almost every data table obtained from merging data from various sources. Finally,
reliable energy data set is prepared after the comprehensive data cleaning in the Energy
Information Administration’s OpenData database.
3.2 Feature Correlation Analysis

The following factors were extracted for use as input features in the analysis. They
were chosen due to their high correlation with electricity price. Numbers indicate
correlation of feature with Average Monthly Retail Electricity Price:
• Natural Gas Consumed, Electric Power Sector (.85)
• Net Electricity Generation (.81)
• Natural Gas Consumed by Industrial Sector (.74)
• GDP (.7)
• Renewables (.77)
The interesting findings was related to the high correlation of the consumption and
production of renewable energy features to the retail price of electricity including solar,
geothermal, wind, and biofuel and basically overall renewable generation were heavily
correlated with the retail price of electricity. Our findings seems to agree that there
exists a relationship explaining the influence of the price of electricity on the adoption
of renewables. In summary, the price of electricity highly depends on price of fuel, the
power import and export to specific geographical area as well as undeniable role of
renewable energy in the network. There are also other major economical factors which
could potentially impact the electricity price such as gross domestic product (Fig. 2).
Fig. 1. The highly correlated points are shown by yellow points and least correlated features are
determined by the dark blue points.
The Fig. 3 demonstrates the correlation of the electricity price with the variation in
natural gas price in California. As it is shown by the feature selection algorithm as well,
the natural gas price has a positive correlation with the electricity price.
Fig. 2. Demonstrates the correlation value and the applied color coding
After preprocessing, S-ARIMA was used to synthetically augment the data from
the 11 selected features from 2001–2014 in order to create projected values to compare
with the 2014–2017 test set. Upon performing linear regression on the augmented
dataset, we obtained MSE = 0.015, MAE = 0.096 and RMSE = 0.125, training
accuracy as 84.9% and test accuracy as 64% which suggested some degree of over-
fitting. Narrowing down to the 6 most seasonal features and training several other
models yielded drastically improved results. The Fig. 4 demonstrates the forecast price
for the PG&E utility in California.
Fig. 3. Demonstrates the variation of the PG&E electricity price with the gas price variation in
California
Fig. 4. Demonstrates the predicted price versus actual price in the PG&E utility are for duration
of three years.
4 Conclusion
Using the time series data method described in this research, a dynamic model capable
of changing weights and adding or adjusting the influence of features over a particular
time bracket in the future, and providing several possible outcomes based on qualitative
parameters like the probability of beneficial or adversarial political or economic events
occurring at a given time is developed. Conceivably, this model would utilize online
learning from databases updated in real time providing constant feedback into the
pipeline and would be able to translate real world economic changes to a corresponding
change in the feature set and weights used in the model. This would allow the user to
consider different scenarios by changing the contributing factors of the prediction.
References
1. Agrawal, R.K., Muchahary, F., Tripathi, M.M.: Long term load forecasting with hourly
predictions based on long-short-term-memory networks. In: 2018 IEEE Texas Power and
Energy Conference (TPEC), pp. 1–6 (2018)
2. Helseth, A., Haugen, M., Jaehnert, S., Mo, B., Farahmand, H., Naversen, C.: Multi-market
price forecasting in hydro-thermal power systems. In: 15th International Conference on the
European Energy Market (EEM), pp. 1–5 (2018)
3. Jiang, L. Hu, G.: Day-ahead price forecasting for electricity market using long-short term
memory recurrent neural network. In: 15th International Conference on Control, Automa-
tion, Robotics and Vision (ICARCV), pp. 949–954 (2018)
4. Varshney, H., Sharma, A., Kumar, R.: A hybrid approach to price forecasting incorporating
exogenous variables for a day ahead electricity Market. In: 2016 IEEE 1st International
Conference on Power Electronics, Intelligent Control and Energy Systems (ICPEICES),
pp. 1–6 (2016)
5. Rafiei, M., Niknam, T., Khooban, M.: Probabilistic forecasting of hourly electricity price by
generalization of ELM for usage in improved wavelet neural network. IEEE Trans. Industr.
Inf. 13(1), 71–79 (2017)
6. Kobayashi, K.: Optimal real-time pricing of electricity based on switched markov chain
models and welfare maximization. In: IEEE 4th International Conference on Cyber-Physical
Systems, Networks, and Applications (CPSNA), pp. 35–39 (2016)
7. Mosbah, H., El-hawary, M.: Hourly electricity price forecasting for the next month using
multilayer neural network. Can. J. Electr. Comput. Eng. 39(4), 283–291 (2016)
8. Ozozen, A., Kayakutlu, G., Ketterer, M., Kayalica, O.: A combined seasonal ARIMA and
ANN model for improved results in electricity spot price forecasting: Case study in Turkey.
In: 2016 Portland International Conference on Management of Engineering and Technology
(PICMET), pp. 2681–2690 (2016)
9. Zhi-Wei, Q.: Mutivariable mutual information based feature selection for electricity price
forecasting. In: 2012 International Conference on Machine Learning and Cybernetics, vol. 1,
pp. 168–173 (2012)
10. Abedinia, O., Amjady, N., Zareipour, H.: A new feature selection technique for load and
price forecast of electrical power systems. IEEE Trans. Power Syst. 32(1), 62–74 (2017)
11. Hehui, Q., Zhiwei, Q.: Feature selection using C4.5 algorithm for electricity price prediction.
In: 2014 International Conference on Machine Learning and Cybernetics, vol. 1, pp. 175–
180 (2014)
12. Wang, K., et al.: A survey on energy internet: architecture, approach, and emerging
technologies. IEEE Syst. J. 12(3), 2403–2416 (2018)
13. Jiang, H., Wang, K., Wang, Y., Gao, M., Zhang, Y.: Energy big data: a survey. IEEE Access
4, 3844–3861 (2016)
14. California Government Open Data Portal (2017). https://data.ca.gov/
15. Python Data Analysis Library (2017)
16. Brownlee, J.: Autoregression Models for Time Series Forecasting With Python – Machine
Learning Mastery (2018)
Queue Formation Augmented with Particle
Swarm Optimisation to Improve Waiting Time
in Airport Security Screening
Mohamad Naji1(&), Ahmed Al-Ani1, Ali Braytee1, Ali Anaissi2,

and Paul Kennedy1
1
University of Technology Sydney, 15 Broadway,
Ultimo, NSW 2007, Australia
mohamad.naji@student.uts.edu.au
2
The University of Sydney, Camperdown, NSW 2006, Australia
Abstract. Airport security screening processes are essential to ensure the safety
of both passengers and the aviation industry. Security at airports has improved
noticeably in recent years through the utilisation of state-of-the-art technologies
and highly trained security officers. However, maintaining a high level of
security can be costly to operate and implement. It may also lead to delays for
passengers and airlines. This paper proposes a novel queue formation method
based on a queueing theory model augmented with a particle swarm optimisa-
tion method known as QQT-PSO to improve the average waiting time in airport
security areas. Extensive experiments were conducted using real-world datasets
collected from Sydney airport. Compared to the existing system, our method
significantly reduces the average waiting time and operating cost by 11.89%
compared to the one-queue formation.
Keywords: Airport security screening process Particle swarm optimisation

Queueing theory Queue formation
1 Introduction and Related Work
The aviation industry continues to grow every year and has become an essential mode
of passenger transportation. In airports all around the world, there has been a contin-
uous increase in the number of passengers. In 2012, for example, 37 million passengers
travelled through Sydney airport, and by 2033 this number is estimated to increase to
74 million passengers. It is also estimated that by 2033, the number of passengers
travelling by air worldwide will increase to 6 billion [1]. In this context, enhancing the
security of the aviation industry and passengers has become essential to ensure the
safety and comfort of passengers when travelling by air. For this reason, researchers
have proposed several methodologies such as queueing theory and fuzzy reasoning to
analyse, optimise and study the effects of congestion caused by security screening
processes [1]. And further to strike a balance between increasing security and reducing
processing time, cost and congestion [2].

https://doi.org/10.1007/978-3-030-15035-8_91
924 M. Naji et al.
Queueing theory, first used in 1970, has mainly been applied to optimise the
screening process of a single server. For example, [2] stated that in order to keep the
system stable, the arrival rate must be less than the service rate. [3] aimed to optimise
the security process. They proposed optimal static and dynamic policies for multi-level
security systems. [4] introduced a b (rejection parameter) to determine whether a bag
requires secondary scanning. [5] applied Parkinson’s Law to reduce service time. [6]
introduced a security and accuracy operating characteristic (SAOC). [7] sought to
improve the security screening process by introducing a selectee lane and assigning to
it passengers that may represent a threat.
Different authors used fuzzy reasoning as a method of optimising the security
screening process, as in the work of [8–10], in which linguistic variables and a fuzzy
inference system were used to analyse technical and human factors and enhance a
baggage screening process. The author in [11] presented a fuzzy reasoning system to
improve the efficiency of a baggage screening system with respect to the uncertainty of
machines (X-ray) and human factors.
Most of the aforementioned studies of airport security screening processes that use
a queueing theory as a method have implemented approaches based on a single server,
such as M/M/1 and M/G/1 (M/M/1 was introduced earlier and in M/G/1, where M
stands for a system with Poisson arrivals, G is generally identically distributed, while 1
is the number of servers [12]). This assumption cannot be practically applied for many
reasons, such as differences in human experience, knowledge of the security screening
process and the existence of special needs passengers who require more processing
time, passengers with slow movement and families with more passengers, as well as
differences between machines such as X-ray machines and metal detectors. We con-
sidered these issues in our model during the implementation phase in order to deter-
mine the average waiting time for a whole system with multiple servers. We consider
an M/M/S/K queueing system with Poisson input, exponential service time, multiple
servers and the buffer or area size (where M represents a system with Poisson arrivals
and exponentially distributed service time, S is the number of servers, and K is the
buffer size).
To the best of our knowledge, this study represents the first attempt to explore the
impact of queue formation on optimising the average waiting time.
The aim of this paper is to use a queueing theory as a method to optimise the
security screening process with multi-servers operating in parallel to serve a different
number of passengers during different seasons, such as Christmas, Easter and school
holidays, and time of the day, as this strongly influences the number of passengers. The
implementation of queueing theory requires certain parameters to be predetermined.
Particle Swarm Optimisation (PSO) is used here to optimise the walking time
parameter at different numbers of servers.
The main contributions of this study are summarised as follows:
• We construct a new queue formation model named QQT-PSO based on the Lindley
process (queueing theory) to improve the average waiting time for multi-servers
operating in parallel for security screening areas.
• We incorporate a new “walking time” parameter in our QQT-PSO model to
improve prediction accuracy.
Queue Formation Augmented with Particle Swarm Optimisation 925
• We model human and server variations to suit real-world application.

• We extensively evaluate the performance of our model on real data collected from
Sydney international airport.
The rest of this paper is structured as follows. Section 2 gives a brief review of
queueing theory and PSO. Section 3 illustrates the model and its implementation.
Section 4 presents the results and discussion, followed by the conclusion.
2 Review of Queueing Theory and Particle Swarm

Optimisation
Different methods are used in this proposed application to optimise the average waiting
time for the security screening process. The following subsection reviews these
methods.
2.1 Brief Review of Queueing Theory

Queueing theory is an approach based on mathematical formalisation and probability
estimates. It has also been used in other real-life queuing applications in contexts such
as airports, trains, hospitals, banks and supermarkets. Queuing theory aims to analyse
and determine the length of queues, service time, average waiting and sojourn time, as
well as other system characteristics. Depending on the application, some queueing
theory values must be predetermined, such as average arrival rate (denoted by k),
average service rate (denoted by l), service and arrival type, number of servers or
channels, buffer size and queue discipline [13].
The behaviour of a queueing system was discussed by Kendall in 1953, who
described a queueing model that includes the service and arrival type, the number of
servers, buffer capacity and service discipline, denoted as follows: A/B/C/D/E, where A
is the inter-arrival time distribution, B is service time distribution, C is number of
servers, D is buffer size, E is the discipline, such as FIFO (first-in-first-out) or LIFO
(last-in-first-out).
In general, for M/M/S, the fraction of time a server is assumed to be busy can be
expressed as:
as
s!ð1sa Þ
Cðs; aÞ ¼ Ps1 ak as
ð1Þ
k¼0 k! þ s!ð1 a Þ
s
while the average waiting and response (sojourn) times respectively are:
Cðs; aÞ l
EðWÞ ¼ ð2Þ
ð1 qÞs
C ðs; aÞ l
E ðRÞ ¼ E ðW Þ þ l ¼ þl ð3Þ
ð1 qÞs
926 M. Naji et al.
The Lindley process defined in Eq. (4) is based on a discrete time stochastic
process, which often used to describe the evolution of a queue length over time or to
determine the average waiting time of passengers in a queue. For example, the first
arriving passenger does not need to wait, so Wn ¼ 0. Subsequent passengers will need
to wait if they arrive at a time before the previous passengers are served [14].
We employed Lindley process in this work due to its simplicity and suitability for
security area queue formation. The simplicity is manifested by the applicability of a
single equation for determining the waiting and sojourn time of passengers.
Wn þ 1 ¼ Wn þ Xn IAn ð4Þ
where Wn is the waiting time of the previous passenger, Xn is the service time of the
previous passenger and IAn is the time between the previous and current arrival (where
both passengers are in service or waiting mode). This IAn parameter gives advantage of
Lindley process over standard queuing theory as presented in Eqs. (1) and (2). Figure 1
illustrates this process given Tn is the arrival time of passenger n and Tn þ 1 is the arrival
time of passenger n þ 1.
Fig. 1. The concept of the Lindley process
2.2 Particle Swarm Optimisation

PSO was developed by Kennedy and Eberhart [15] as a search optimisation algorithm. It
has been successfully applied in many application domains and produced very accurate
results [16, 17]. Accordingly, we decided to incorporate PSO with the proposed queueing
theory model, hence our proposed model is termed QQT-PSO. PSO is used in our model to
optimise and find the best solution for the walking time for the security screening process.
The first step of PSO is to initialise the population by assigning random velocities
and positions. At every step changing the velocity for every particle, so the particles
cooperate to find the best location in the search space to find the optimal solution
towards its Pbest and Gbest: Pbest is the personal or individual best particle and Gbest
is the global best for the best in the population.
The acceleration is weighted by a random term with separate random numbers
being generated for acceleration toward Pbest; Gbest. According to [17], if we have a
particle ðiÞ, the update velocity and position at time ðt þ 1Þ will be as follows:
Vi ðt þ 1Þ ¼ ðwvi ðtÞ þ r1 c1 ðpi ðtÞ xi ðtÞÞ þ r2 c2 ðgðtÞ xi ðtÞÞÞ ð5Þ

x i ð t þ 1Þ ¼ x i ð t Þ þ v i ð t þ 1 Þ ð6Þ
where
xi ðtÞ is the particle position for particle i, vi ðtÞ is the particle velocity for particle i, r1 ; r2
are uniform distribution random numbers between 0 and 1, c1 ; c2 are acceleration
coefficients and w is the inertia coefficient to control the impact of the previous velocity
on the updated velocity.
3 QQT-PSO Method
In our implementation, we take into consideration multiple servers in parallel with

different service distributions, i.e., the M/M/S queueing theory assumption, to be
applicable to real-world airport security settings. The QQT-PSO framework consists of
multiple steps as shown in Fig. 2.
Fig. 2. The QQT-PSO frame work
Step 1: To collect the data from the airport and divide the data into two groups. The
first group is to train the PSO model to optimise the Wwalking for different queue
formations. The second group is used to examine the QQT-PSO model to determine the
optimised average waiting time.
Step 2: The aim of PSO is to optimise the objective function known as the fitness
function. In this setting, we follow a wrapper approach based on PSO to seek the global
best walking time that minimises our fitness function defined in Eq. (9). PSO iteratively
tries to opt for the optimal walking time which produces an identical result to the real
data.
The walking time is evaluated based on our fitness function to opt for the best value
that can lead the fitness function fast to zero. A new population of walking time is
generated in each iteration. For each evaluation, PSO optimises a different number of
walking time sets, which is based on the number of servers used during that time.
However, some parameters in PSO have to be predetermined before optimisation and
this includes the following:
The lower bound decision was set at 3 s (aka varmin), while the upper bound of the
decision variable was set at 25 s (aka varmax). The population size was set at 50 and
the number of iterations was set at for 100.
928 M. Naji et al.
Step 3-QQT-PSO Model: In general, most airports use a single passenger queue for
the security screening process. In this study, however, we consider three queuing
scenarios: single, two and three queues. Also, we investigate their impact on passenger
waiting time. The scenarios of our queue formation are presented as follows:
(i) One-queue scenario: all seventeen servers are allocated to one queue area and the
passenger at the top of the queue is allocated to one of the S parallel servers, as
shown in Fig. 3.
Fig. 3. First queueing scenario: one queue
(ii) Two-queue scenario: The total area A is divided into two equal queues or buffers
and when there is an even number of servers, each area has the same number,
whereas when there is an odd number of servers, the size of each buffer is based
on the number of servers allocated to it. For example, in the case of seventeen
servers, eight servers are allocated to the first queue and nine servers are allo-
cated to the second queue, as shown in Fig. 4.
Fig. 4. Second queueing scenario: two queues

(iii) Three-queue scenario: The total area is divided into three queues. The size of
each buffer is based on the number of servers allocated to it. For example, if
there are seventeen servers, the first two queues are allocated six servers each
and the third queue is allocated five servers, as shown in Fig. 5.
Fig. 5. Third queueing scenario: three queues
Depending on the queue formation, a passenger is assigned by the security officers

to join a certain queue (1, 2 or 3), based on the available free space and based on the
type of passengers, such as normal passengers (aka economy passengers), crew,
business passengers, families or passengers with special needs. Passengers are expected
to be served by one of the S servers according to an exponential distribution with a
mean service rate of l and service completion rate of 1/l, while the capacity of the area
is assumed to be finite.
If N is the total number of passengers in the security area waiting to be served by S
servers, then PN is the steady state probability of N passengers in the system.
Ps
In general, PN Q1 ¼ PN Q1 Si is the total number of passengers served by S
i¼1
servers in the first queue, while the total number of passengers served in the second
queue is PN Q2 , and the total number of passengers served in the system (all queues) is
PN Qtotal ¼ PN Q1 þ PN Q2 þ . . . þ PN QN ð7Þ
The probability of blocking is defined as the chance that a customer will lose
service due to high demand and lack of resources. For example, a probability of
blocking of 0.01 means that 1% of customers will be denied or will lose service. The
probability of blocking for different numbers of queues for M/M/S/Q (where M rep-
resents a system with Poisson arrivals and exponentially distributed service time, S
stands for the number of servers and Q is the number of queues) is

kl q
Q X
X S 1 kl
r ðrÞ
PQ ¼ qþ1
ð8Þ
q¼1 r¼1 ð1 kl
rÞ
930 M. Naji et al.
Equation (8) illustrates that the probability of blocking decreases when the number
of queues increases.
To be more applicable to queueing processes such as those at an airport security
screening area and to produce accurate results, another parameter such as walking time
must be incorporated in Lindley equation. Thus, the modified formula is
Wn þ 1 ¼ Wn þ Xn þ Wwalking IAn ð9Þ
where Wwalking is the time required by a passenger to walk when the passenger comes to
the top of the queue and is the first to be served.
A modified Lindley equation Eq. (9) is used to determine the average waiting time
of N passengers.
According to [4], knowledge of the passenger screening process differs from one
passenger to another. Also, as proposed in [18–20], the efficacy of screening machines
can vary from one to another. Security officers also vary in their experience, which
could affect the time needed to screen luggage (service time). Finally, according to [21]
and in line with the data collected, the time required for a passenger to complete the
security screening process varies between 15 and 60 s for economy passengers (aka
normal lanes), and is 30 s on average, so a time of 30 s is chosen as the mean average
of the exponentially distributed service time. Families and passengers with special
needs (aka slow lanes) take three-and-a-half times longer than normal passengers,
whereas business passengers and the crew (aka express lanes) require half the time
compared to normal passengers. These reasons support our case to apply different
service times for various types of passengers while we optimise the passenger’s
average waiting time. The three service time variations are as follows:
1. Normal lanes: the service time is assumed to be exponentially distributed with
average service mean of 30 s and then multiplied by a random number generator
between 1 and 2. The aim of this variation is to ensure variability among passen-
gers. This is applied to all servers that serve normal passengers, and the service time
of passengers is expressed by the equation:
ðh: lnðRÞÞ ð10Þ
where h is the exponential distributed service time and ðRÞ is random number
generator between 1 and 2.
2. Slow Lanes: various passengers such as families and passengers with special needs
may take three-and-a-half longer than normal passengers. This case is implemented
using the equation:
ððh: lnðRÞÞ 3:5Þ ð11Þ
3. The express or fast lanes: these lanes serve passengers including crew, staff and
business passengers. These passengers may require half the service time of normal
passengers. The fast lane is implemented using the equation:
ððh: lnðRÞÞ 0:5Þ ð12Þ

4 Experiment Results and Discussion

4.1 Data Collection
Our experiments are performed on real data collected from Sydney international airport
(T1). This data was collected from 17 lanes (aka servers). These servers are classified
into three categories: express, normal and slow lanes for special needs passengers and
families. The data were collected at different periods of the year to ensure it incor-
porates all the seasonal variations, including both busy periods (e.g. Christmas and
Easter) and non-busy periods (i.e. January and February).
4.2 Results
The first experiment is to optimise the walking time parameter introduced in Eq. (9) by
applying the PSO algorithm. The inter-arrival rate between two consecutive passengers
is exponentially distributed at the rate of 1/k s. The area size is set to be compatible
with the Sydney airport security area, but the area size cannot be disclosed due to
confidentiality.
In the following sections, we present the results of our QQT-PSO model for one
queue, and then we compare the results with the actual results (airport).
Generally speaking, the aims of this experimental study are twofold. One is to
examine how significant the proposed augmented QQT-PSO model is by producing an
accurate and close result to the airport’s actual system. The second target is to measure
the improvement in the QQT-PSO average waiting time compared with the actual (one
queue) airport result.
4.2.1 PSO Result

The optimised walking time depends on the number of servers operating during that
time, for example, if the number of servers is one then the optimised walking time
using the PSO is one. If the number of servers is seventeen, then the value of the
optimised walking time is seventeen.
For the one-queue formation, the value of the walking time varies from 5.4 s to
15.6 s, depending on the location of the server. However, the values of the walking
time improved by 31.5% and 51.2 for two and three queues, respectively. This is what
we anticipated from reducing passenger’s walking time.
4.2.2 QQT-PSO Result

In the first part of this experiment, we provide a one-queue model which produces a
similar average waiting time compared to the actual airport result.
4.2.2.1 One-Queue Formation

Table 1 show that QQT-PSO produces a high similarity when the numbers of servers
are more than four servers. For example, the similarity varies between 93% and 98.5%
when the number of servers increases from four to seventeen, whereas when the
932 M. Naji et al.
number of servers varies between one and three, the QQT-PSO produces an average
waiting between 66% and 88%. The proposed system is able to produce a standard
error (S) prediction interval of 88.16% similar to the actual values across all of the
servers, also considering the variation in service time between passengers, security
officers and machines.
The mean square error shows the error changes between a range 4.77 and 37.17 s
when there are more than three servers. The mean square error increases from 129.66 to
218.54 s when there are three or less than three servers. However, the paired t-test
between the actual results and QQT-PSO produces a p-value of 0.6894, but this result
is not significant at p 0.05. Hence, these results suggest that our result is very
similar to the airport result.
Table 1. Average waiting time and standard error (QT-PSO (1Q, 2Q and 3Q) and actual result)
No. of servers AWT (seconds) MSE (seconds2) S (second) AWT (seconds)
QQT-PSO
Real QQT-PSO Actual-1Q Actual-1Q (2Q) (3Q)
17 389.002 397.406 29.140 5.398 354.307 330.214
16 174.557 180.167 21.183 4.602 163.267 152.168
15 170.535 175.397 19.710 4.439 158.402 147.947
14 256.651 253.112 13.557 3.682 233.384 219.38
13 201.054 196.695 17.742 4.212 178.665 168.302
12 161.045 165.578 18.089 4.253 149.861 141.169
11 160.100 159.545 7.206 2.684 147.229 139.131
10 159.125 162.772 12.980 3.602 150.123 139.131
9 179.370 179.626 4.773 2.184 168.897 160.452
8 150.931 148.600 10.983 3.314 140.415 133.675
7 146.261 140.907 18.994 4.358 135.057 128.709
6 120.291 116.852 14.481 3.805 110.439 105.469
5 98.314 93.915 20.275 4.502 90.810 87.086
4 135.742 129.267 37.174 6.097 128.389 125.821
3 84.228 74.721 129.668 11.387 74.721 74.721
2 77.983 65.972 162.123 14.564 65.972 65.972
1 26.285 39.775 218.549 17.565 39.774 39.774
Average 158.322 165.033 95.933 5.920 146.453 138.730
Figure 6 shows the average waiting time for the proposed model compared to the
actual result from the airport. This figure shows that the QQT-PSO produces very
accurate results compared to the actual result and it also follows the same trend.
Fig. 6. Average waiting time (QT-PSO (1Q, 2Q and 3Q)-actual airport result)
4.2.2.2 Two-Queue Formation

Figure 6 shows and Table 1 shows the results for one, two and three queues and the
actual result for the average waiting time per server and the average for the total
system.
From Table 1, we can deduce the following:
When we compare server to server, it can be seen that the enhancement in the
average waiting time varies from 1 s when there are less than 5 servers to approxi-
mately 43.099 s when there are 17 servers. Moreover, the total average improvement
per server is approximately 11 s, which is a good achievement. However, the total
system’s average waiting time reduced from 2680.31 to 2489.71 s. Also, the two-queue
formation improves the average waiting time from 157.66 to 146.45 s. Figure 6 shows
the two-queue formation still follows the same trend and when the number of servers
increases, the average waiting time decreases dramatically. For example, when the
number of servers increases from eight to seventeen, the average waiting time
decreases to a larger degree.
4.2.2.3 Three-Queue Formation

The three-queue formation experiment includes two comparisons. The first comparison
is when we change the area from two queues to three queues and the second com-
parison is when we change the security area from one to three queues to investigate the
impact on average waiting time. Table 1 shows the results.
1. Two-Three Queue Formation:
The minimum enhancement in average waiting time occurs when there is a small
number of servers. For example, when the number of servers varies between three
and nine, the enhancement in average waiting time varies from 2.568 to 8.445 s.
However, when the number of servers is between ten and seventeen, the
enhancement in average waiting time increases to 24.80 s. The average improve-
ment in average waiting time is 7.68 s per server.
934 M. Naji et al.
2. One-Three Queue Formation:

The minimum enhancement in average waiting time occurs when there is a small
number of servers. For example, when the number of servers varies between three
and nine, the enhancement in average waiting time varies from 3.44 to 19.15 s.
However, when the number of servers is between ten and seventeen, the
enhancement in average waiting time increases to 58.35 s. The average improve-
ment in average waiting time is 18.93 s per server.
Figure 6 plots the average waiting time for one, two and three queues and the actual
result and shows the improvement in average waiting time when the area changes from
one queue to two queues. The average waiting time is further enhanced when the area
changes to three queues from two queues. For example, the change from one to two
queues improves the average waiting time by 7.11%. However, the change from two to
three queues improves average waiting time by 4.78%. So, the total system enhance-
ment in terms of average waiting when the security screening area is changed from one
to three queues is 11.89%.
The results of the proposed method are compared with the actual results from
Sydney airport. The outcome of the experiments shows that the queue formation
methodology improves the average waiting time for the airport security screening
process. Also, the walking time factor has an extensive impact on average waiting time.
Finally, we didn’t compare the results of the proposed method with any of the state-of-
the-art, because to the best of our knowledge, this study represents the first attempt to
explore the impact of queue formation on optimising average waiting time.
5 Conclusion
Airport security screening is essential to ensure the safety of the aviation industry and
passengers travelling by air. Experienced officers and advanced technological instru-
ments help to enhance security. However, the delay caused by security screening
remains a concern. Technical methods have been proposed to model and optimise this
process. Despite the achieved improvements, delays still present a major concern. This
paper used queueing theory to build a system with multi-lane parallel servers to study
the impact of queue formation on the average service time. Unlike existing methods
which do not consider service time variations between the different servers, our pro-
posed method successfully takes this issue into consideration. The results show the
proposed method extensively improved the system average waiting time by 11.98%.
Further studies are required to develop new methods which address the limitations of
the existing methods and the theory behind them by taking into consideration human
and machine factors to determine the optimal numbers of servers required to serve a
different number of passengers to achieve an acceptable waiting time.
References
1. Naji, M., et al.: Airport security screening process: a review. In: CICTP 2017.
ASCE LIBRARY (2017)
2. Gilliam, R.R.: An application of queueing theory to airport passenger security screening.
Interfaces 9(4), 117–123 (1979)
3. Lee, A.J., Jacobson, S.H.: The impact of aviation checkpoint queues on optimizing security
screening effectiveness. Reliab. Eng. Syst. Saf. 96(8), 900–911 (2011)
4. Babu, V.L.L., Batta, R., Lin, L.: Passenger grouping under constant threat probability in an
airport security system. Eur. J. Oper. Res. 168(2), 633–644 (2006)
5. Marin, C.V., et al.: Human factors contributes to queuing theory: Parkinson’s law and
security screening. In: Proceedings of the Human Factors and Ergonomics Society Annual
Meeting. SAGE Publications (2007)
6. Olapiriyakul, S., Das, S.: Design and analysis of a two-stage security screening and
inspection system. J. Air Transp. Manage. 13(2), 67–74 (2007)
7. Nie, X., et al.: Simulation-based selectee lane queueing design for passenger checkpoint
screening. Eur. J. Oper. Res. 219(1), 146–155 (2012)
8. Skorupski, J., Uchroński, P.: A fuzzy model for evaluating airport security screeners’ work.
J. Air Transp. Manage. 48, 42–51 (2015)
9. Skorupski, J., Uchroński, P.: Fuzzy inference system for the efficiency assessment of hold
baggage security control at the airport. Saf. Sci. 79, 314–323 (2015)
10. Skorupski, J., Uchroński, P.: A fuzzy system to support the configuration of baggage
screening devices at an airport. Expert Syst. Appl. 44, 114–125 (2016)
11. Skorupski, J., Uchroński, P.: A fuzzy reasoning system for evaluating the efficiency of cabin
baggage screening at airports. Transp. Res. Part C: Emerg. Technol. 54, 157–175 (2015)
12. Cooper, R.B.: Introduction to Queueing Theory. North Holland (1981)
13. Avi-Itzhak, B., Levy, H., Raz, D.: Quantifying fairness in queueing systems: principles and
applications. Preprint (2004)
14. Asmussen, S.: Applied Probability and Queues, vol. 51. Springer, Heidelberg (2008)
15. Kennedy, J., Eberhart, R.: Particle swarm optimization. In: Proceedings of IEEE
International Conference on Neural Networks, vol. IV (1995)
16. Braytee, A., et al.: ABC-sampling for balancing imbalanced datasets based on artificial bee
colony algorithm. In: 2015 IEEE 14th International Conference on Machine Learning and
Applications (ICMLA). IEEE (2015)
17. Del Valle, Y., et al.: Particle swarm optimization: basic concepts, variants and applications in
power systems. IEEE Trans. Evol. Comput. 12(2), 171–195 (2008)
18. Wang, X., Zhuang, J.: Balancing congestion and security in the presence of strategic
applicants with private information. Eur. J. Oper. Res. 212(1), 100–111 (2011)
19. Martín-Cejas, R.R.: Tourism service quality begins at the airport. Tour. Manage. 27(5), 874–
877 (2006)
20. Almazroui, S., Wang, W., Zhang, G.: Imaging technologies in aviation security. Adv. Image
Video Process. 3(4), 12 (2015)
21. Kirschenbaum, A.A.: The cost of airport security: the passenger dilemma. J. Air
Transp. Manage. 30, 39–45 (2013)
Polar Topographic Derivatives for 3D
Face Recognition: Application to Internet
of Things Security
Farshid Hajati1(B) , Ali Cheraghian2 , Omid Ameri Sianaki1 , Behnam Zeinali3 ,

and Soheila Gheisari4
1
College of Engineering and Science, Victoria University Sydney, Sydney, Australia
{farshid.hajati,omid.amerisianaki}@vu.edu.au
2
College of Engineering and Computer Science,
The Australian National University, Canberra, Australia
u5942646@anu.edu.au
3
Electrical Engineering Department,
Iran University of Science and Technology, Tehran, Iran
zeinali.behnam@gmail.com
4
Faculty of Engineering and Information Technology,
University of Technology Sydney, Ultimo, Australia
soheila.gheisari@student.uts.edu.au
Abstract. We propose Polar Topographic Derivatives (PTD) to fuse the

shape and texture information of a facial surface for 3D face recognition.
Polar Average Absolute Deviations (PAADs) of the Gabor topography
maps are extracted as features. High-order polar derivative patterns are
obtained by encoding texture variations in a polar neighborhood. By
using the and Bosphorus 3D face database, our method shows that it is
robust to expression and pose variations comparing to existing state-of-
the-art benchmark approaches.
1 Introduction
Automatic face recognition has many commercial, military and security appli-
cations such as Internet of Things (IoT) [1,2]. Despite the significant progress in
face recognition technology, recognizing a face under the unconstrained condition
is still challenging. There are five major challenging problems in face recogni-
tion which are changing the illumination condition, facial expressions, head pose
variations, occlusions, and aging [3]. Many of the face recognition algorithms
are categorized as holistic which use the whole face image for feature extraction.
Alyuz et al. [7] constructed seven meaningful regions: the nose, the left/right eye,
the forehead, the left/right cheek, and the mouth-chin. By partitioning the 3D
face image into different paths their scheme was fast and accurate. To heighten
the accuracy, the results were fused in decision level after applying the Lin-
ear Discriminate Analysis (LDA) to patches. Smeets et al. [6] proposed a local
feature method for 3D face, called meshSIFT in which important points are
https://doi.org/10.1007/978-3-030-15035-8_92
Polar Topographic Derivatives for 3D Face Recognition 937
detected as extrema in a scale space then a canonical orientation is assigned to

those points based on the surface normals in the scale-dependent local neighbor-
hood. Concatenated histograms of slant angles and shape indices is described
as a feature vector. In Recent methods in the face recognition have used both
the texture (2D) and the shape (3D) information to improve the performance.
Bowyer et al. research [5] the fact that multimodal face recognition systems have
better accuracy compared to the shape and the texture-based face recognition
algorithms was investigated. Chang et al. [4] showed a new method by apply-
ing a PCA-based algorithm to the both texture and shape of face image and
fusing the matching scores. Using a gallery of 275 subjects, a recognition rate
of 93% and 99% for 3D and multimodal face recognition was reported respec-
tively. Malassiotis and Strintzis [10] shows a method to overcome pose variations
in the corresponding texture map by using the range map. In Bronstein et al.
research [11], by finding an isometric-invariant representation of the face, the
effects of expressions were removed and new expression-invariant multimodal
face recognition algorithm was proposed.
In Mian et al. [12] proposed method, in order to achieve efficiency and robust-
ness to facial expressions, a fully automatic multimodal face recognition algo-
rithm and performs hybrid (feature based and holistic) matching was used.
The novel 3D Spherical Face Representation (SFR) and Scale-Invariant Fea-
ture Transform (SIFT) descriptor were used to form a rejection classifier and
the results of all matching engines were fused at the metric level to achieve
higher accuracy. Hajati et al. [8] proposed new algorithm in which expression
and pose invariant 3D face recognition features were extracted by using Patch
Geodesic Distance (PGD). In PGD, 3D face images are partitioned into equal-
sized square patches in a non-overlapping manner. To encode the shape adjusted
textures into patterns, Local geodesic paths within patches and global geodesic
paths for patches are combined.
2 Polar Topographic Derivatives (PTD)
The topography map is one of the most important shape representation methods
in 3D space. A topography map shows the physical features of a surface and
their structural relationships. A topography map reveals the relief of a surface
by means of level curves (contour lines). Lines that joint equal-height points on
the surface above or below a reference point is known as a Level curves. The
highest point in the topography map usually known as a reference point. Figure 1
indicates an example of level curves in a face surface and the corresponding
topography map.
Imagine a surface z(x, y), the level curve with the height index h, Lh , is
obtained by
Lh = {z(x, y)|zmax − hwz < z(x, y) < zmax − (h − 1)wz };

h = 1, 2, ..., zmax /wz + 1 (1)
938 F. Hajati et al.
(a) (b) (c)
Fig. 1. Level curves and topography map. (a) original 3D face surface, (b) surface level
curves, (c) topography map
where zmax is the maximum height (reference) point in the surface, wz is the step
size of the level curve, h is the height index of the level curve, and · denotes
the floor function.
Given the level curves h , the topography map T (x, y) is indicated as

1, if (x, y, z) ∈ h
T (x, y) = (2)
0, / h
if (x, y, z) ∈
After extracting the topography map from the face surface, Gabor wavelet is
applied to the topography map. The Gabor wavelets are represented as follows
[9]:

kμ,v
2
kμ,v
2
z2 σ2
ϕμ,v (z) = 2 exp − [exp (ikμ,v ) − exp(− )] (3)
σ 2σ 2 2
where μ and v denote the orientation and the scale of the Gabor wavelets,
respectively. In Eq. (3), kμ,v is defined as
kμ,v = kv eiϕμ (4)
where kv = kmax /f v and ϕμ = πμ/8. kmax is the maximum frequency, and f is

the spacing factor between kernels in the frequency domain. We use Gabor ker-
nels with five different scales, v ∈ {0, ..., 4}, and eight√orientations, μ ∈ {0, ..., 7},
with the parameters σ = 2π, kmax = π/2, and f = 2.
Polar Gabor topography maps are the convolution of the topography map
with the Gabor kernels in the polar coordinates as
P GTμ,v (r, θ) = T (x, y) ∗ ϕμ,v (z) (5)

2 2
r= (x − xrp ) + (y − yrp ) (6)
⎧
⎪
⎨ 0, if x = xrp , y = yrp
−1 y−yrp
θ= tan ( x−xrp ), if x ≥ xrp (7)
⎪
⎩ tan−1 ( y−yrp ) + π,
x−xrp if x < xrp
where xrp and yrp are the coordinates of the reference point (nose tip). Polar
Gabor topography maps are divided into non-overlapping polar patches. Here,
we label the patch points with p and q indexes as
p = r/wr + 1; 0 ≤ r < rmax (8)
q = θ/wθ + 1; 0 ≤ θ < 2π (9)
the patch sizes in the radial and the angular directions are shown by wr and wθ
respectively, and indicate as
wr = rmax /Nr (10)
wθ = 2π/3(2p − 1) (11)
where rmax and Nr indicate the maximum radius and the number of sectors in
the radial direction, respectively.
In the proposed polar patching, the total amount of patches, Np , is
rmax /wr

Np = 3(2k − 1) (12)
k=1
The (p, q)th patch of the Gabor topography map, GTpq (rpq , θpq ), is repre-
sented as
pq
P GTμ,v (rpq , θpq ) = P GTμ,v (wr (p − 1) + rpq , wθ (q − 1) + θpq ) (13)
where rpq and θpq are the coordinates of the (p, q)th patch. For each patch,
we compute Polar Average Absolute Deviation (PAADs) from the mean of the
patch’s gray levels. For the (p, q)th patch of the polar Gabor topography map,
the PAAD feature from the mean of the gray levels is defined as

pq 1

pq pq pq pq pq pq

P AADμ,v = P GTμ,v (r , θ ) − P GTμ,v (r , θ ) (14)
N rpq pq
θ
pq pq pq
where P GTμ,v (r , θ ) is the (p, q)th patch of the Polar Gabor topography map.
pq
N is the number of pixels in the (p, q)th patch and P GTμ,v (rpq , θpq ) is the mean
of the (p, q)th patch’s gray levels.
By concatenating the PAADs of (p, q)th patch in different scales and orien-
tations, we obtain a new feature vector, called Polar Gabor Topography Vector
(PGTV), as

P GT V pq = P AADμ,v pq
|μ = 0, . . . , 7 & v = 0, . . . , 4 (15)
pq
where P AADμ,v is computed using the μ and v as the scale and the orientation.
We also use high-order polar information in the texture map of the object to
extract discriminant features. First, we transform the texture map into the polar
coordinates. Given a texture map, f (x, y), the transformed texture map in the
polar coordinates, F (r, θ), is defined as
F (r, θ) = f (x, y) (16)
where r and θ are computed using Eqs. (6) and (7).

We partitioned the texture map into predefined patches and extracted deriva-
tives within patches to compute local features. We partitioned the texture map
into non-overlapping polar patches using Eqs. (8)-(11) as
F pq (rpq , θpq ) = F (wr (p − 1) + rpq , wθ (q − 1) + θpq ) (17)
where rpq and θpq are the coordinates of the (p, q)th patch of the partitioned
texture map.
Let F pq (r0pq , θ0pq ) be a point of the (p, q)th patch of the texture map, the
eight neighbors of the point are defined as
F pq (ripq , θjpq ) = F pq (rpq + i.Δr, θpq + j.Δθ) ; i, j = −1, 0, 1 & i = j = 0 (18)
where Δr and Δθ are the radial and the angular resolutions of the texture map.
Here, we assume Δθ = π · Δr/180.
The angular directions with different orders and the derivative along the
radial can be computed after determining the 8-neighborhood of the point
F pq (r0pq , θ0pq ). The first-order polar derivatives along the radial and the angu-
lar directions of F pq (r0pq , θ0pq ) are defined as
∂F pq (r0pq , θ0pq ) /∂r = F pq (r0pq , θ0pq ) − F pq (r1pq , θ0pq ) (19)
∂F pq (r0pq , θ0pq ) /∂θ = F pq (r0pq , θ0pq ) − F pq (r0pq , θ1pq ) (20)
Similarly, the second-order polar derivatives are defined as
∂ 2 F pq (r0pq , θ0pq ) /∂r2 = ∂F pq (r0pq , θ0pq ) /∂r − ∂F pq (r1pq , θ0pq )/∂r (21)
∂ 2 F pq (r0pq , θ0pq ) /∂θ2 = ∂F pq (r0pq , θ0pq ) /∂θ − ∂F pq (r0pq , θ1pq )/∂θ (22)
Generally, the nth -order polar derivatives are defined using the (n−1)th -order
polar derivatives as
∂ (n) F pq (r0pq , θ0pq ) /∂r(n) =

∂ (n−1) F pq (r0pq , θ0pq ) /∂r(n−1) − ∂ (n−1) F pq (r1pq , θ0pq )/∂r(n−1) (23)
∂ (n) F pq (r0pq , θ0pq ) /∂θ(n) =

∂ (n−1) F pq (r0pq , θ0pq ) /∂θ(n−1) − ∂ (n−1) F pq (r0pq , θ1pq )/∂θ(n−1) (24)
By using the unit step function, the computed derivatives are encoded
into binary bits, there for binary patterns called Polar Derivative Patterns
(PDP) are formed. The nth -order Polar Derivative Patterns (PDP) along
the radial direction, P DPr (F pq (r0pq , θ0pq )), and along the angular direction,
P DPθ (F pq (r0pq , θ0pq )), in the (p, q)th patch are defined as
P DPrpq (F pq (r0pq , θ0pq )) =
20 .u ∂ (n) F pq (r0pq , θ0pq )/∂r(n) · ∂ (n) F pq (r1pq , θ1pq )/∂r(n) + · · · (25)
+27 .u ∂ (n) F pq (r0pq , θ0pq )/∂r(n) · ∂ (n) F pq (r0pq , θ1pq )/∂r(n)
P DPθpq (F pq (r0pq , θ0pq )) =

20 .u ∂ (n) F pq (r0pq , θ0pq )/∂θ(n) · ∂ (n) F pq (r1pq , θ1pq )/∂θ(n) + · · · (26)
+27 .u ∂ (n) F pq (r0pq , θ0pq )/∂θ(n) · ∂ (n) F pq (r0pq , θ1pq )/∂θ(n)
where u(·) is the unit step function determines the direction of the derivatives
transition.
It can be easily observed that by increasing the operator order from the
first-order to the second-order, the third-order, and up to the fourth-order, finer
details are extracted from the face surface. It is also evident that in higher
order patterns, e.g. fourth-order, more noise comes into the features rather than
identity information.
Eventually, by concatenating the spatial histogram to make a feature vector
as the distribution of the PDPs is modeled:
HP DP pq = {HP DPrpq , HP DPθpq } (27)
where HP DPrpq and HP DPθpq
are the PDP’s histogram vectors extracted in the
radial and the angular directions in the (p, q)th patch, respectively.
For measure the dissimilarity between the query image and the model, two
distance vectors are computed. The first distance vector is Polar Gabor Topog-
raphy Distance Vector (PGTDV) which demonstrates the difference between the
two patches in the query topography map f and the model topography map g:
P GT DV pq = {[P GT V pq ]f − [P GT V pq ]g } (28)
The second distance vector is Polar Derivative Distance Vector (PDDV)
which shows the distance between the patches in the query texture map f and
the model texture map g:
P DDV pq = {[HP DP pq ]f − [HP DP pq ]g } (29)
By using P GT DV pq , and P DDV pq , we can compute dissimilarity (D)
between the query image and the model as
rmax /wr 2π/wθ rmax /wr 2π/wθ

D = ((1−wb ).( P DDV pq 2 )+(wb ).( P GT DV pq 2 ))1/2
p=1 q=1 p=1 q=1
(30)
where the symbol
·
represents the Euclidean norm of the vector and wb is a
balancing weight determined experimentally.
The robustness of the proposed algorithm in the face recognition task under the
pose rotation and the expression variation is investigated. In our experiments,
the Bosphorus database [14] is used that contains faces with pose rotations
and expression changes. In the experiments, preprocessing was applied to face
scans in the database. Spikes in 3D faces were removed using median filtering in
the z-coordinate. Holes were filled using cubic interpolation and 3D scans were
resampled on a uniform square grid at 1 mm resolution. Finally, all scans were
subjected to pose normalization and cropped to the size of 120 × 120 pixels.
The Bosphorus database [14] contains 105 subjects (61 females and 44 males)
with a total of 4666 faces. The images include frontal views with different
facial expressions, pose variations, and occlusions. The resolution in x, y, and
z (height) dimensions are 0.3 mm, 0.3 mm, and 0.4 mm respectively, and color
texture images are high resolution (1600×1200 pixels). To compute the accuracy
of the proposed method, we consider faces with expression and pose variations
in the database (see Fig. 2). The robustness against the pose rotation is one
of the most major problems in the accuracy computing of the face recognition
systems. Here, we evaluate the acceptable performance of the proposed method
using faces with pose variations from the Bosphorus database [14]. There are 11
different poses including left 45◦ , right 10◦ , right 20◦ , right 30◦ , right 45◦ , right
down, right up, slightly up, strong up, slightly down, and strong down for each
subject. Since our goal is to solve the problem of the face recognition using one
exemplar per person case, as the gallery, one frontal image per enrolled subject
is used.
The rank-1 recognition rate of the our approach is shown in Table 1. We
compare proposed method with the recognition rates of the methods in [8] and
[10] as the benchmarks. For faces with small yaw rotations (right 10◦ , right
20◦ , and right 30◦ ), the rank-1 recognition rate is higher than 98%, while it is
higher than 91% for faces with large yaw rotations (right 45◦ and left 45◦ ). In
upward pitch rotations, our method achieves 100% recognition rate for small and
Fig. 2. Images of a subject in the Bosphorus database [14] with different poses and
expressions.
Table 1. Comparison the rank-1 recognition rates of the proposed method and the
benchmarks under the pose variations.
Considered Proposed method Hajati et al. [8] Malassiotis and

rotation Strintzis method [10]
Right 10o 100 92.3 65.7
o
Right 20 100 88.6 52.4
Right 30o 98.1 80 42.9
o
Right 45 91.4 39 9.5
Left 45o 91.4 38.1 21
Up (Slight) 100 87.6 81.9
Up (Strong) 100 79 69.5
Down (Slight) 99.1 87.6 84.8
Down (Strong) 97.1 69.5 68.6
Right-Up 94.3 63.8 31.4
Right-Down 90.5 34.3 21.9
Average 96.5 69.1 50
strong rotations, while the benchmarks have 87.6% and 81.9% for small rota-
tions and 79% and 69.5% for strong rotations, respectively. For downward pitch
rotations, the accuracy rate of the proposed method is 99.1% and 97.1% in small
and strong rotations, respectively. The rank-1 recognition rate of the proposed
method is 94.3% and 90.5% for faces with the right-upward rotation and the
right-downward rotation, respectively. Overally, the proposed method achieves
the average recognition rate of 96.5% across all rotations in the database, out-
performing [8] and [10] with 69.1% and 50%, respectively.
We also evaluate the performance of the proposed method under the expres-
sion changes using the Bosphorus database [14]. For this purpose, we report the
rank-1 recognition rate of the proposed method using all faces with pose and
Table 2. Comparison the rank-1 recognition rates of the proposed method and the
benchmarks under the expression variations.
Method Recog. Rate Gallery Probe

Baseline ICP [15] 72.4% 47 (Frontals) 1508 (Expressions)
Baseline PCA [15] 70.6% 47 (Frontals) 1508 (Expressions)
Alyüz et al. [16] 95.3% 47 (Frontals) 1508 (Expressions)
Dibeklioğlu et al. [17] 89.2% 47 (Frontals) 1527 (Expressions)
Hajati et al. [8] 69.1% 105 (Frontals) 1155 (Expressions)
Proposed Method 96.6% 105 (Frontals) 3696 (Expressions +
Rotations)
expression variations as the probe (see Table 2). The rank-1 recognition rate for
the proposed method is 96.6% using 105 galleries and 3696 probes, while the
best one among the benchmarks is 95.3% using 47 galleries and 1508 probes.
4 Conclusion
In this paper, a new approach called Polar Topographic Derivatives (PTD) was
proposed for 3D face recognition. The PTD used both the range and the tex-
ture maps to extract discriminative features. We applied the PTD to expression
and pose invariant face recognition purpose, and achieved considerable improve-
ments in the recognition rates compared to the state-of-the-art methods. The
experimental results demonstrate that the algorithm has an acceptable perfor-
mance even though in large database. Although the proposed algorithm have
been applied to 3D face recognition task, it can be generalized to other textured
3D shapes.
References
1. Hajati, F., Tavakolian, M., Gheisari, S., Gao, Y., Mian, A.S.: Dynamic texture
comparison using derivative sparse representation: application to video-based face
recognition. IEEE Trans. Hum.-Mach. Syst. 47(6), 970–982 (2017)
2. Hajati, F., Cheraghian, A., Gheisari, S., Gao, Y., Mian, A.S.: Surface geodesic
pattern for 3D deformable texture matching. Pattern Recognit. 62, 21–32 (2017)
3. Abate, A.F., Nappi, M., Riccio, D., Sabatino, G.: 2d and 3d face recognition: a
survey. Pattern Recognit. Lett. 28(14), 1885–1906 (2007)
4. Chang, K.I., Bowyer, K.W., Flynn, P.J.: Face recognition using 2d and 3d facial
data. In: ACM Workshop on Multimodal User Authentication, pp. 25–32 (2003)
5. Bowyer, K.W., Chang, K., Flynn, P.: A survey of approaches and challenges in 3d
and multi-modal 3d + 2d face recognition. Comput. Vis. Image Underst. 101(1),
1–15 (2006)
6. Smeets, D., Keustermans, J., Vandermeulen, D., Suetens, P.: meshsift: Local sur-
face features for 3d face recognition under expression variations and partial data.
Comput. Vis. Image Underst. 117(2), 158–169 (2013)
7. Alyuz, N., Gokberk, B., Akarun, L.: Regional registration for expression resistant
3-d face recognition. IEEE Trans Forensics Secur. 5(3), 425–440 (2010)
8. Hajati, F., Raie, A.A., Gao, Y.: 2.5d face recognition using patch geodesic
moments. Pattern Recognit. 45(3), 969–982 (2012)
9. Lee, T.: Image representation using 2d gabor wavelets. IEEE Trans Pattern Anal.
Mach. Intell. 18(10), 959–971 (1996)
10. Malassiotis, S., Strintzis, M.G.: Robust face recognition using 2d and 3d data: pose
and illumination compensation. Pattern Recognit. 38(12), 2537–2548 (2005)
11. Bronstein, A.M., Bronstein, M.M., Kimmel, R.: Three-dimensional face recogni-
tion. Int. J. Comput. Vis. 64(1), 5–30 (2005)
12. Mian, A.S., Bennamoun, M., Owens, R.: An efficient multimodal 2d–3d hybrid
approach to automatic face recognition. IEEE Trans. Pattern Anal. Mach. Intell.
29(11), 1927–1943 (2007)
13. Zhang, H., Gao, W., Chen, X., Zhao, D.: Learning informative features for spatial
histogram-based object detection. In: Proceedings of the International Conference
on Neural Networks, vol. 3, pp. 1806–1811 (2005)
14. Savran, A., Alyüz, N., Dibeklioğlu, H., Çeliktutan, O., Gökberk, B., Sankur, B.,
Akarun, L.: Biometrics and identity management. Bosphorus Database for 3D Face
Analysis, pp. 47–56 (2008)
15. Alyüz, N., Gökberk, B., Dibeklioğlu, H., Savran, A., Salah, A., Akarun, L., Sankur,
B.: Biometrics and identity management. In: Biometrics and Identity Management,
pp. 57–66. Springer, Heidelberg (2008)
16. Alyüz, N., Gökberk, B., Akarun, L.: A 3d face recognition system for expression and
occlusion invariance. In: Proceedings of the IEEE Second International Conference
on Biometrics Theory, Applications and Systems, pp. 1–7 (2008)
17. Dibeklioğlu, H., Gökberk, B., Akarun, L.: Nasal region-based 3d face recognition
under pose and expression variations. In: Proceedings of the Third International
Conference on Advances in Biometrics, vol. 5558, pp. 309–318 (2009)
A Survey on Conversational Agents/Chatbots
Classification and Design Techniques
Shafquat Hussain(&), Omid Ameri Sianaki, and Nedal Ababneh
College of Engineering and Science, Victoria University Sydney,

Sydney, Australia
{shafquat.hussain,omid.amerisianaki,
nedal.ababneh}@vu.edu.au
Abstract. A chatbot can be defined as a computer program, designed to interact

with users using natural language or text in a way that the user thinks he is
having dialogue with a human. Most of the chatbots utilise the algorithms of
artificial intelligence (AI) in order to generate required response. Earlier chatbots
merely created an illusion of intelligence by employing much simpler pattern
matching and string processing design techniques for their interaction with users
using rule-based and generative-based models. However, with the emergence of
new technologies more intelligent systems have emerged using complex
knowledge-based models. This paper aims to discuss chatbots classification,
their design techniques used in earlier and modern chatbots and how the two
main categories of chatbots handle conversation context.
Keywords: Chatbots Conversational agents Classifications

Design techniques conversational context Algorithms Rule based
Retrieval based Generative based Chatscript AIML Pattern matching
Machine learning Neural network
1 Introduction
The interaction between humans and computer systems is shifting towards natural
language-based interfaces. Natural language has recently been viewed as a compelling
enabling technology for personalization, which allows each user to interact with the
system in his or her own words, rather than using one of a small number of preset ways
to interact [1]. A text based, or a voice-based conversation is usually initiated by the
user who asks a question using natural language and the chatbot answers the question
using natural language [2]. Chatbots are primarily designed to conduct a conversation
with humans using natural language dialogues via auditory or textual methods, in such
a way that humans can easily communicate with the chatbots.
Chatbots have already come a long way from their roots in systems that were more
about simple chat or fun and entertainment [2]. Although modern chatbots use the
power of artificial intelligence to answer users’ complex questions but current state-of-
the art systems are still a long way from being able to have coherent, contextual and
natural conversations with humans [3].

https://doi.org/10.1007/978-3-030-15035-8_93
A Survey on Conversational Agents/Chatbots Classification and Design Techniques 947
Chatbot technology is not new; it was started in 1960’s when Joseph Weizen-
baum’s program ELIZA, published in 1966, which seemed to be able to fool users into
believing that they were conversing with a real human [4]. Early chatbots like ELIZA
used keyword matching and minimal context identification and lacked the ability to
keep a conversation going. ELIZA program inspects the user input for keywords and
once a keyword is found, it transformed the sentence according to the rules associated
with that keyword in a script. Response is generated by assembly/reassembly rules
associated with selected decomposition rules. Keywords and their associated trans-
formation rules constitute the SCRIPT (it is data and not a part of the program itself) for
a particular class of conversation [5]. A.L.I.C.E. (Artificial Linguistic Internet Com-
puter Entity) is another natural language processing chatbot and is one of the strongest
programs of its type. It has won the Loebner Prize, awarded to accomplished huma-
noid, talking robots, three times (in 2000, 2001 and 2004) [6]. A.L.I.C.E. is built using
Artificial Intelligence Mark-up Language (AIML), based on eXtensible Markup Lan-
guage (XML), developed over the past 10 years [7]. AIML supports most of the
modern chatbot platforms in use today. AIML based chatbot are the most popular
because they are lightweight and easy to configure. Predefined responses for a given set
of questions also induces constraint on the conversational agent, in the sense that
appropriate context also needs to be considered while generating these responses.
Developers and researchers continue to add new components with improved func-
tionalities to the existing methodologies and in some cases, have even proposed a new
design architecture. Most of these improvements were largely based on utilization of
ontologies and recollecting details from the current and previous conversation [8].
Chatbots are generally designed either to serve some specific task, also called task-
oriented chatbots or to a simple chit-chat, also called non-task oriented chatbots, for fun
and entertainment. In this paper, the broad and main classification of chatbots is pre-
sented first which is followed by the exploration of popular chatbot design approaches
used in the categories mentioned above in this paper.
2 Classification of Chatbots
Over the last few years, the chatbot field has become so dynamic with the arrival of new
technologies that a precise classification of chatbots has become subjective to the scope
of their use. Chatbots could be classified into various categories based on several criteria,
e.g. mode of interaction, knowledge domain, their usage and the design techniques
(response generation method) that are typically employed in building these chatbots.
These criteria may include the core design philosophy of the chatbots or the extent to
which context needs to be stored and considered in understanding the conversation or the
type and purpose of the conversation for which the chatbot needs to be designed [8]. The
broad classification can be done based on the following criteria (Fig. 1);
1. Interaction Mode (Text based, or voice/speech based)
2. Chatbot Application (Task-Oriented or Non-Task-Oriented)
3. Rule-based or AI (Machine learning, Deep learning etc.)
4. Domain-Specific or Open-Domain
948 S. Hussain et al.
Fig. 1. Broad classification of Chatbots
But in general, chatbots are classified into two main categories based on goals
[9–11] (Fig. 2):
1. Task-oriented chatbots
2. Non-task-oriented chatbots
Fig. 2. Main classification of Chatbots
Task-oriented chatbots are designed for a particular task and are set up to have short
conversations, usually within a closed domain. Unlike task-oriented chatbots, non-task
oriented chatbots can simulate a conversation with a person and seem to perform chit-
chat for entertainment purpose in open domains [10].
3 Chatbots Design Techniques
There are several different approaches that can be used to develop a task-oriented or
non-task oriented chatbot. These approaches can overlap between these two main
categories of chatbots. This section explores those approaches, in a broad scope, used
in the development of chatbots since the very early chatbot ELIZA. These approaches
can be divided into three main categories [8, 10, 12–15];
1. Rule-Based Approaches
2. Retrieval-Based Approaches
3. Generative-Based Approaches
Each of these main approaches can employ following multiple techniques;
1. Parsing
2. Pattern Matching
3. AIML
4. Chatscript
5. Ontologies
6. Markov Chain Model
7. Artificial Neural Networks Models
7:1 Recurrent Neural Networks (RNNs)
7:2 Sequence to Sequence Neural Model
7:3 Long Short-Term Memory Networks (LSTMs)
1. Parsing. Parsing is a method that takes the text as an input and extract meaningful
information from it by converting that text into a set of more uncomplicated words
(lexical parsing) that can be easily stored and manipulated. It determines the
grammatical structure of a sentence. A simplistic parsing technique was used in an
early chatbot, ELIZA to parse the input text for the keyword in the sentence. The
keyword is then matched against the documents in the corpus to find the appropriate
response to the user [16]. A more advanced parsing technique is semantic parsing,
which converts user’s input sentence to a machine-understandable representation of
its meaning. It is used in some of the modern commercial chatbots tools like
Dialogflow to understand the user’s intent [17].
2. Pattern Matching. Pattern matching is the most commonly used approach in chat-
bots to classify the user input as <pattern> and produce a suitable response stored
in <template>. These <pattern> <template> pairs are hand crafted. Although pattern
matching techniques are used in both early and modern chatbots, the complexity of the
algorithms used in them differs. Early chatbots like ELIZA used simple pattern
matching rules, as compared to A.L.I.C.E, which uses more complex pattern matching
rules and associates some degree of conversational context while searching the stored
categories for a matching <pattern> [7]. Pattern matching technique is mostly used in
QA chatbots. This techniques has advantage of being flexible to create conversations.
The approach has some disadvantages in terms of scaling as all the possible patterns
are built manually and it’s not a trivial task. Due to the scaling issue, chatbots infor-
mation extraction capabilities are limited, and its responses could be predictable and
repetitive.
3. AIML. AIML (Artificial Intelligence Mark-up Language) has been derived from
XML (Extensible Mark-up Language). AIML is designed to create conversational
flow in chatbots. AIML is an open standard. AIML is made up of data objects which
are called AIML objects. These AIML objects, which are also called AIML
elements, are made up of units which are called topics and categories. The topic is
an optional top-level element that has a name attribute and a set of categories related
to that topic. Categories are the most fundamental unit of knowledge in an AIML
file. A category has a minimum of two more elements in it called ‘pattern’ and
‘template’. Each category is a rule for matching an input and converting it to an
output. The ‘pattern’ matches against the user input and the ‘template’ is used to
generate the Chatbot’s answer [18]. AIML is powerful in designing chatbots’
conversational flow, flexible and easy to use but needs some Natural language
processing and programming expertise.
4. Chatscript. Chatscript is an open source authoring tool for building chatbots.
Chatscript is a combination of natural language engine and dialog management
system designed for interactive conversation while maintaining user state across the
conversations [19]. It is a rule-based engine where rules are created in program scripts
through a process called dialog flow scripting. These scripts can be stored as a normal
text file. Machine learning tools can also be used to mine users conversations logs to
improve your dialog flows. Chatscript make use of concepts. A concept of all nouns
or adverbs can be created. An existing database of some 2000 predefined concepts
and scripters can also easily write their own concepts [20].
5. Ontologies. Domain ontologies are used in chatbots to replace hand crafted domain
knowledge with ontological domain knowledge. Use of ontologies is not new as
they have been used within specific dialogue system modules, for example in
language generation as the basis of the systemic grammar approach [21]. The main
advantage of using domain ontology in a chatbot is that the chatbot can explore the
concepts nodes of an ontology to establish the relationship between concepts that
are used in conversation but can also imply new reasoning [22]. However more
general use of ontologies in chatbots is relatively small.
6. Markov Chain Model. A Markov chain model is a probabilistic model and the
rules in the Markov chain process are based on probabilities. The Markov chain
seeks to model probabilities of state transitions over time. The key idea behind
Markov chain model is that given the current state, there is a fixed probability for
one or more states that it will go to next. When a chatbot uses this method, it will
produce an output that goes with the state transition. This allows the chatbot to
construct sentences for responses that are more suitable probabilistically whilst
being different every time but are more or less coherent. The starting state can be
based on the user input, giving the response some relevance. Markov Chains are a
popular method to create chatbots for entertainment purposes that imitate simple
human conversation. Because Markov chain model is a simplified version of a
complex decision-making process, it does not work well to emulate rich and
complex conversation [8]. A simple Markov chains model is easy to program and
the whole model can be summarized in a matrix.
7. Artificial Neural Networks Models. The latest advances in machine learning,
specifically in artificial neural networks, has made it possible to develop more
intelligent chatbots. Artificial Neural Networks based chatbots can use both
retrieval-based or generative-based approaches in terms of response generation,
although the research trend is moving into the generative-based approach [23, 24].
The main difference between rule-based and artificial neural network based chatbot
development approaches is the presence of a learning algorithm in the artificial

neural network approach. Artificial neural networks models are one of the learning
algorithms used within machine learning [25]. There are supervised and unsuper-
vised machine learning algorithms. Deep learning, which is a subset of machine
learning, is capable of learning unsupervised from unstructured or unlabeled data.
Deep learning imitates the human brain function in terms of processing data and
creating patterns that are used for decision making. Artificial neural networks have
been used on a variety of tasks, including computer vision, decision making, speech
recognition, machine translation, social network filtering, and medical diagnosis,
etc. Many different variants of artificial neural networks are used for natural lan-
guage processing [26]. The use of deep learning neural networks has risen in the
field of conversational modelling, especially the recurrent neural network (RNN),
sequence to sequence and long short term memory networks (LSTMs) have dom-
inated the field [24].
7:1 Recurrent Neural Networks (RNNs). A Recurrent Neural Network (RNN) is
a class of artificial neural network and a variant of a recursive artificial neural
network. It works on the principle of saving the output of a layer and feeding
this saved output to the new input in order to predict the next output. In other
words, a recurrent neural network has ability to remember the previous com-
putations and uses this understanding of previous information in current pro-
cessing. This simple approach changed the chatbots landscape. A RNN can be
visualized as multiple copies of a neural network with the output of one serving
as an input to the next [27]. Traditional neural networks do not possess this
ability to analyze the history of information. Because of its ability to remember
previous computations, recurrent neural networks are useful when you need to
process sequential data. RNN has the ability to capture the inherent sequential
nature in natural language where words in the natural language develop their
semantical meaning based on the previous words in the sentence. This allows
the RNN to retain context and produce an output based on the previous words
in the sentence. This approach makes recurrent neural network suitable for
chatbots as understanding the conversational context is essential to understand
the users inputs and produce a contextually correct responses [28]. In chatbots,
if they are fed with the information from the previous conversations, they tend
to produce more contextually correct responses.
7:2 Sequence to Sequence (Seq2Seq) Neural Models. Seq2Seq model is based on
recurrent neural network (RNN) architecture and consists of two recurrent
neural networks, an encoder that processes the input and a decoder that pro-
duces the output [28]. Both encoder and decoder allow the model to be fed with
variable length input sentences. It was first proposed in 2014 as a variation of
Ritter’s generative model that uses advances in deep learning to achieve higher
accuracy. The encoder encodes the <status> or input sentence, and the decoder
decodes the <status> and generates the desired <response>. The model is pri-
marily used in language translation (e.g. for statistical machine translation)
where the input sentence is in one language and the <response> is a translated
sentence in another language. This model can also be applied to chatbots to
convert between input <status> and output <response> [29]. The Sequence to
Sequence model (Seq2Seq) is the current industry best practice for response
generation and is widely used [29].
7:3 Long Short-Term Memory Networks (LSTMs). Long Short-Term memory
networks (LSTMs) are a special kind of recurrent neural networks [8]. LSTMs
are designed to avoid the long term dependency problem of RNNs. Memory
cells and gates are introduced in LSTMs which allows the cells to remember the
previous information for long periods of time. These memory cells can store
information in, write new information and read information from them like data
in a computer’s memory. The flow of information is controlled with the use of
input gates, forget gates and output gates [30]. An LSTM network is ideal for
learning from experience as compared to a traditional RNN. LSTM networks
have since replaced RNNs as standard for this task. A well-trained LSTM
network has the capability to perform better classification, processing and
prediction of time series even when there is a long period of gaps of unknown
size between significant events. These capabilities exemplify LSTM’s high
performance as compared to other available RNNs, hidden Markov models, and
other sequence learning methods that are being used in many applications. Thus
LSTMs prove to be highly useful in designing chatbots due to their ability to
refer to a piece of distant information in time very frequently [8, 30].
4 Task-Oriented Chatbots
The main aim for task-oriented chatbots are to help the user to achieve a certain task.
They are designed for dealing with specific scenarios such as: booking a hotel/flight,
booking accommodations, placing an order for a product, scheduling an event, or
helping users to access some specific information etc. Personal assistants such as
Cortana, Alexa and Siri are examples of voice based task-oriented chatbots/
conversational agents who attempt to return an answer for the task they receive.
Task-oriented chatbots work well in restricted domains. These chatbots do not possess
general knowledge and you cannot ask them trivia questions. Rather, they are goal
oriented chatbots focused on helping you achieve a specific goal.
5 Non-task-Oriented Chatbots
Non-task oriented chatbots are systems designed for extended conversations, set up to
mimic the unstructured conversational or ‘chats’ characteristic of human-human
interaction, rather than focusing on a particular task like booking plane flights. These
systems often have an entertainment value. Non-task oriented chatbot architectures fall
into two classes:
1. Generative-Based such as sequence-to-sequence models, which generate proper
responses during the conversation
2. Retrieval-Based which learn to select responses from the current conversation from
a repository.
Generative models can generate more proper responses that could have never
appeared in the corpus. Corpus-based systems mine large datasets of human-human
conversations, which can be done by using information retrieval (IR-based systems
simply copy a human’s response from a previous conversation) or by using a machine
translation paradigm such as neural network sequence-to- sequence systems, to learn to
map from a user utterance to a system response. Retrieval-based models enjoy the
advantage of informative and fluent responses [31], because they select a proper
response for the current conversation from a repository with response selection algo-
rithms. A variety of heuristics can be applied in order to choose a proper response. It
may be based on the fairly simple concept of a rule-based expression match or may
well be as complex as using a combination of Machine Learning classifiers. Systems
based on this concept do not produce new responses but just choose one from a pool of
predefined responses. Rule-based systems include the early influential ELIZA and
PARRY systems.
6 Dialogue Context in Task-Oriented Chatbots
As demonstrated earlier, the typical objective of a task-oriented chatbot is to perform

certain tasks, such as ordering a pizza, scheduling an event or troubleshooting an
internal problem within the device [32]. Normally, this type of chatbot is utilised in
devices in order to provide convenience to the users. The dialogue context in this case
is either in the form of text conversation or may be in the form of natural language. In
both cases, a different algorithm is followed by the chatbots. In case of a text-based
conversation, an architecture of pattern matches is typically used. The chatbots are
programmed with the feature of pattern matching in order to group the text. Based on
this, an appropriate response is generated for the users. For instance, a single task can
be written differently by different users [33]. In that case, the chatbot must be able to
understand the context of the conversation to proceed to the particular task. As an
example, two different sentences, including “Order one large vegetable pizza” and “I
want to have one large veggie pizza” may have similar meaning for the user; however,
when this is processed with a typical chatbot without any functionality of pattern
matches, the response might not be according to the expectations of the users [34]. The
feature of pattern matching enables the bots to understand the context and based on
that, an appropriate response is generated.
7 Dialogue Context in Non-task-Oriented Chatbots
Non-task oriented chatbots typically do not have any goal to achieve and are not
designed to handle any particular task. Nonetheless, these chatbots are very helpful in
different applications, such as keeping elderly people company and improving the
understanding of the learners of the second language. In each of its applications, unlike
task-oriented chatbots, they might not have to observe a set of similar tasks; rather, they
would be required to respond to the users in order to proceed on the discussion [35].
Therefore, the algorithm designing in case of non-task oriented chatbots will be
complex as compared to a task-oriented chatbot. In some cases, these chatbots are also
combined with task-oriented chatbots in order to provide an enhanced experience to the
users for a smooth transition. The dialogue context in this case will be random and the
user might not have discussion related to any particular context [36]. For instance, the
user may ask the chatbot for information related to weather, or the question might be
related to horoscope or any recent news story. Therefore, the chatbot must be capable
of retrieving information from online resources in order to communicate effectively
with the user.
8 Conclusion
In this paper, we have explored chatbot’s broad and general classification and discussed
various design techniques used to design chatbots. To conclude, it can be argued that
the typical interaction between human and computer systems has been modernised and
natural languages have become a major input in contemporary chatbots. Natural lan-
guage has recently been viewed as a compelling enabling technology for personal-
ization, which allows each user to interact with the system in his or her own words,
rather than using one of a small number of preset ways to interact. In handling such
queries based on the natural language, the chatbots are programmed using different
strategies and algorithms in order to generate an effective response for the user. In
earlier discussion on the topics, the chatbots can be typically divided into two main
categories including task-oriented chatbots and non-task oriented chatbots. As the name
highlights, the task-oriented chatbots are designed to perform particular tasks based on
the instructions of the user while a non-task oriented chatbot will have several purposes
but will not be able to perform any specific task. The dialogue context in both of these
chatbots will be different. For instance, in case of a task-oriented chatbot, it might have
to process a set of similar inputs and simple algorithms based on pattern matching can
be used to process the queries. In case of non-task oriented chatbot, the queries may
vary in terms of their context and several algorithms and strategies would be required to
use in order to assure correct response of the system. Generative-based chatbots are
difficult to build and operate as compared to retrieval-based chatbots.
References
1. Zadrozny, W., et al.: Natural language dialogue for personalized interaction. Commun. ACM
43(8), 116–120 (2000)
2. Hussain, S., Athula, G.: Extending a conventional chatbot knowledge base to external
knowledge source and introducing user based sessions for diabetes education. In: 2018 32nd
International Conference on Advanced Information Networking and Applications Work-
shops (WAINA) 2018. IEEE
3. Levesque, H.J.: Common sense, the turing test, and the quest for real AI: Reflections on
Natural and Artificial Intelligence. MIT Press, Cambridge (2017)
4. Wikipedia: Chatbot. 29 December 2018 (cited 30 December 2018). https://en.wikipedia.org/
wiki/Chatbot
5. Weizenbaum, J.: ELIZA–a computer program for the study of natural language commu-
nication between man and machine. Commun. ACM 26(1), 23–28 (1983)
6. Wikipedia. Artificial Linguistic Internet Computer Entity. 19 November 2018 (cited 30
December 2018). https://en.wikipedia.org/wiki/Artificial_Linguistic_Internet_Computer_
Entity
7. Wallace, R.S.: The anatomy of ALICE. In: Parsing the Turing Test, pp. 181–210. Springer,
Dordrecht (2009)
8. Ramesh, K., et al.: A Survey of Design Techniques for Conversational Agents. In:
International Conference on Information, Communication and Computing Technology.
Springer, Singapore (2017)
9. Budulan, S.: Chatbot Categories and Their Limitations (2018). (cited May 2018). https://
dzone.com/articles/chatbots-categories-and-their-limitations-1
10. Chen, H., et al.: A Survey on Dialogue Systems: Recent Advances and New Frontiers
(2017). arXiv preprint arXiv:1711.01731
11. Yan, Z., et al.: Building Task-Oriented Dialogue Systems for Online Shopping. In: AAAI
(2017)
12. Masche, J., Le, N.-T.: A Review of Technologies for Conversational Systems. In:
International Conference on Computer Science, Applied Mathematics and Applications.
13. Abdul-Kader, S.A., Woods, J.: Survey on chatbot design techniques in speech conversation
systems. Int. J. Adv. Comput. Sci. Appl. 6(7), 72–80 (2015)
14. Mathur, V., Singh, A.: The Rapidly Changing Landscape of Conversational Agents (2018).
arXiv preprint arXiv:1803.08419
15. Bradeško, L., Mladenić, D.: A survey of chatbot systems through a loebner prize
competition. In: Proceedings of Slovenian Language Technologies Society Eighth Confer-
ence of Language Technologies (2012)
16. Weizenbaum, J.: ELIZA—a computer program for the study of natural language
communication between man and machine. Commun. ACM 9(1), 36–45 (1966)
17. Clarke, D.: Three AI Technologies that could make chatbots intelligent. 24 March 2018
(cited 18 December 2018). https://hackernoon.com/three-ai-technologies-that-could-make-
chatbots-intelligent-10f8c6e8b4b0
18. Shawar, B.A., Atwell, E.: Chatbots: are they really useful? Zeitschrift für Computerlinguistik
und Sprachtechnologie, p. 29 (2007)
19. Wikipedia contributors. ChatScript. Wikipedia, The Free Encyclopedia. 4 September 2018,
19:19 UTC. https://en.wikipedia.org/w/index.php?title=ChatScript&oldid=858055386. Acce-
ssed 22 Jan 2019
20. Robino, G.: How to build your first chatbot using Chatscript (2018). (cited 18 December
2018). https://medium.freecodecamp.org/chatscript-for-beginners-chatbots-developers-c58b
b591da8
21. Milward, D., Beveridge, M.: Ontology-based dialogue systems. In: Proceedings of 3rd
Workshop on Knowledge and reasoning in practical dialogue systems (IJCAI 2003) 2003
22. Al-Zubaide, H., Issa, A.A.: Ontbot: Ontology based chatbot. In: 2011 Fourth International
Symposium on Innovation in Information & Communication Technology (ISIICT). IEEE
(2011)
23. Nuez Ezquerra, A.: Implementing ChatBots using Neural Machine Translation Techniques
(2018). Universitat Politècnica de Catalunya
24. Csáky, R.: Deep Learning Based Chatbot Models (2017). https://doi.org/10.13140/rg.2.2.
21857.40801
25. Pelk, H.: Machine learning, neural networks and algorithms (2016). (cited 18 January 2019).
https://itnext.io/machine-learning-neural-networks-and-algorithms-4cd27a105b89
26. Wikipedia contributors. Artificial neural network. In: Wikipedia, The Free Encyclopedia.
https://en.wikipedia.org/w/index.php?title=Artificial_neural_network&oldid=878429976.
Accessed 13:24, 25 Jan 2019
27. Young, T., et al.: Recent trends in deep learning based natural language processing. IEEE
Comput. Intell. Mag. 13(3), 55–75 (2018)
28. Cahn, J.: CHATBOT: Architecture, design, and development. In: University of Pennsylvania
School of Engineering and Applied Science Department of Computer and Information
Science (2017)
29. Cho, K., et al.: Learning phrase representations using RNN encoder-decoder for statistical
machine translation (2014). arXiv preprint arXiv:1406.1078
30. Sak, H., Senior, A., Beaufays, F.: Long short-term memory recurrent neural network
architectures for large scale acoustic modeling. In: Fifteenth Annual Conference of the
International Speech Communication Association (2014)
31. Ji, Z., et al.: An information retrieval approach to short text conversation (2014). arXiv
preprint. arXiv:1408.6988.1
32. Duijst, D., Sandberg, J., Buzzo, D.: Can we Improve the User Experience of Chatbots with
Personalisation? (2017)
33. Liu, B., Xu, Z., Sun, C., Wang, B., Wang, X., Wong, D.F., Zhang, M.: Content-oriented user
modeling for personalized response ranking in Chatbots. IEEE/ACM Trans. Audio Speech
Lang. Process. (2018)
34. Pilato, G., Augello, A., Gaglio, S.: A modular architecture for adaptive ChatBots. In:
Proceedings of 5th IEEE International Conference on Semantic Computing, ICSC 2011
(2011)
35. Vtyurina, A., Savenkov, D., Agichtein, E., Clarke, C.L.A.: Exploring conversational search
with humans, assistants, and wizards. In: Proceedings of the 2017 CHI Conference Extended
Abstracts on Human Factors in Computing Systems – CHI EA 2017 (2017)
36. Niculescu, A.I., Banchs, R.E.: Strategies to cope with errors in human-machine spoken
interactions: using chatbots as back-off mechanism for task-oriented dialogues. In: Errors by
Humans Mach. multimedia, multimodal Multiling, Data Process (2015)
Dimensionality Reduction for Network
Anomalies Detection: A Deep
Learning Approach
Ahmed Dawoud(&), Seyed Shahristani, and Chun Raun
School of Computing, Engineering, and Mathematics,

Western Sydney University, Sydney, Australia
{a.dawoud,s.Shahristani,c.raun}@westernsydney.edu.au
Abstract. Cyber threats are a severed challenge in current communications

networks. Several security measures were introduced to at different network
layers to enhance security. One of the common networking security solutions is
intrusion detection and prevention systems, with more focus on detecting the
attacks. Various approaches are being used in network threat detection, for
instance, signature-based and anomalies detection methods. Signature-based
depends on a database of predefined attacks signature, in operation, the systems
compare the traffic against the signature, if a match occurs, then an attack is
identified. This approach cannot detect attacks that do not have a signature in the
database. The anomalies detection approach utilizing various approaches to
define the threats, for instance, statistical, and machine learning algorithms.
Several machine learning algorithms had been used for network anomalies
detection. A major common deficiency was poor accuracy, which kept the
approach not industrially applicable. In this paper, we propose a framework for
network anomalies detection. The proposed framework showed improvement in
detection accuracy. The framework adopts semi-unsupervised algorithms for
novelty detection to tackle the rapid development in the cyber security attacks.
The framework embraces the unsupervised deep learning in more elegant
technique, where it dramatically reduces the features from the first phase.
1 Introduction
The recent expansion of communications networks such as the Internet of Things,

cloud computing, and software-defined networks, these technologies introduced new
security challenges and threats. Security solutions, e.g. firewalls, and Intrusion
Detection System (IDS), have been deployed as a standard security measure to detect
and prevent threats at different networking layers. However, the recent security chal-
lenges require innovative security solutions.
IDS is hardware or a software system dedicated to revealing security threats. It
utilizes several approaches to detection. i.e. signature-based and anomalies detection.
Signature-based systems experience deficiencies in detecting unprecedented attacks,
while the attacks signatures do not exist in its system profile. Anomalies detection
overcome this deficiency. However, the cost is results are inexact with a high rate of
false true and positive false alarms [1].

https://doi.org/10.1007/978-3-030-15035-8_94
958 A. Dawoud et al.
Anomalies based detection systems utilize various methodologies, for example,

statistical, and artificial intelligence. A recognizable artificial intelligence approach is
machine learning algorithms, e.g. Bayes networks, Supported Vector machines SVM,
and several variations of Principal Components Analysis PCA. However, the accuracy
was not industrially applicable. Likewise, shallow neural networks did not prevail with
exceptional accuracy [1–3].
A decade ago, breakthroughs in machine learning algorithms primarily, achieve-
ments in deep neural networks, revived the research interests in using neural networks
in several domains [4–6]. Deep learning neural network is a neural network, with
multiple hidden layers between the input and output layers. The first layer represents
the inputs (data features), while the last layer is the network output. Deep neural
networks existed for a long time [3], however; it was not possible to train the network
for several reasons, including the vanishing gradient decent in backpropagation, poor
generalization, and intensive computation power required. The recent advances in DL
started in 2006 by a pre-training step using restricted Boltzmann Machines (RBM) [4].
Subsequently, several algorithms were proposed to solve generalization problem like
using rectifier Linear Units (ReLU) instead of sigmoid and using dropouts [6].
The paper studies the potentials of applying DL as a pertaining phase to reduce the
dimensionality. The dimensionality reduction is an essential step to improve the
detection accuracy for network anomalies detection. The research focuses on unsu-
pervised algorithms, as they prone to detect new threats. The study focuses on
autoencoders. We propose a detection technique, where the framework consists of two
phases. The first phase based on Unsupervised Deep Learning Algorithm UDLA for
dimensionality reduction. While in the second stage, the output from the first phase is
forwarded to a simple clustering algorithm e.g. K-Means.
The rest of this paper structured as follow. The second section is a background and
related work on deep learning and anomalies detection; it represents a literature review
for ongoing research on DL and network anomalies detection. In the third section, we
discuss the dimensionality reduction as a design principal for the framework.
Section four represents the proposed detection framework; in the fifth section, we
provide implementation and experimental evaluation of the framework.
2 Deep Learning for Anomalies Detection
The deep learning has been used successfully for several applications. In image
recognition, Hinton et al. utilized Deep Belief Network (DNN) on MNIST a numbers
handwritten images dataset, the algorithm recorded 1.25% error rate while the second
lowest was 1.4% scored by SVM [4]. In 2012, ImageNet competition the convolutional
neural network decreased the error rate of 16.4%; the dataset included 14 m labelled
images and 1k categories [6]. In voice recognition. Mohamed et al. utilized DBN to
reduce the phoneme error rate from 26% to 20.7% in TIMIT dataset for voice
recognition [7].
An inclusive study of unsupervised machine learning algorithms evaluated seven
algorithms [8]. The study used classic KDD99 dataset. It deduced all algorithms per-
formed poorly in remote code execution escalation attacks. SVM and Y-means scored
Dimensionality Reduction for Network Anomalies Detection 959
reasonable accuracy in detecting privileges escalation attacks. Additionally, C-means

recorded the worst accuracy compared to other algorithms.
The concept of dimensionality reduction refers to projecting highly dimensioned
data on a lower subspace without a significant loss of data meaning. Where, in a lower
dimensional data, the discrimination between normal and abnormal classes is evident.
Principal Components Analysis PCA is a dimensionality reduction algorithm learns
linear relationships. Kernel PCA is a non-linear version of the basic PCA. To represent
nonlinear relationships, a kernel function is used to map the data to higher dimensions,
and then the PCA reduces the dimensionality. Autoencoders algorithms imply
dimensionality reduction concepts, as it converts the data into a new representation that
keep most significant features (encoder), and then execute a reconstruction phase
(decoder). Various studies compared Autoencoders as a dimensionality reduction
algorithm with PCA and its nonlinear extension kernel PCA. An experiment introduces
a comparison between PCA, KPCA, Autoencoders, and Autoencoders shows a sig-
nificant performance of Autoencoders [8]. The study applied the four algorithms on the
artificial dataset and two real datasets.
Fiore et al. deployed a semi-supervised deep learning technique to detect anomalies
in networks traffic [10]. Their approach was base on a Restricted Boltzmann Machines
(RBM). The algorithm showed poor accuracy, specifically, when testing with new
samples. Various research papers focused on integrating other machine learning
approaches and deep learning algorithms. Mostafa et al. employed Deep Belief Net-
work DBN as an unsupervised phase, the pass the output to SVM in the second phase
[11]. They claimed their approach approximately scored 93% accuracy where each
algorithm individually achieved 88% for DBN and 90% for SVM. In a broader
comparative study on anomalies detection, the authors presented a deep structured
energy-based model. The study compares their algorithms in two different decisions
boundaries against five classical anomalies detection algorithms including PCA and
SVM. The authors went one further step by applying their algorithm to various data
types, i.e., static, sequential, and spatial datasets [12]. KDD99 dataset was one of
several data sets they used in their experiments. Their results showed or better accuracy
compared to algorithms like PCA and kernel PCA.
3 Autoencoders for Dimensionality Reduction
The Applying unsupervised algorithm as a pre-training step is common practice in

many frameworks to enhance the accuracy. For example, Fig. 1 shows a classification
framework where it includes a pre-training unsupervised step; then the outputs are
passed to an SVM which do the classification. However, this comes with the cost of
additional computations resources, e.g. memory, processing, and time.
One of the goals of the framework is reducing the additional cost of computations,
directly, rather than adding a series of inputs to the second phase-possibly hundreds or
thousands of features- a singular value is passed.
Fig. 1. Deployment autoencoders as a pre-training phase
The first stage of the suggested framework is based on the idea of dimensionality
reduction. The concept of dimensionality reduction refers to projecting highly
dimensioned data on a lower subspace without a significant loss of data meaning.
Additionally, in a lower dimensional data, the discrimination between normal and
abnormal classes is evident. Principal Components Analysis PCA is a dimensionality
reduction algorithm learns linear relationships. Kernel PCA is a non-linear version of
the basic PCA. To represent nonlinear relationships, a kernel function is used to map
the data to higher dimensions, and then use PCA to reduce the dimensionality.
Autoencoders algorithms imply dimensionality reduction concepts, as it converts
the data into a new representation that keep most significant features (encoder), and
then execute a reconstruction phase (decoder).
Figure 2 depicts the basic structure of an autoencoder includes two symmetric input
and output layers, in between a series of hidden layers with a lower number of features
(neurons), reducing the no of features is a reduction of dimensionality. However, at the
output layers, autoencoders reconstruct the same number of input features.
V1
Fig. 2. Generic auto encoder architecture

The autoencoder implicitly conducts the dimensionality reduction. To use

autoencoders for dimensionality reduction, only the encoder part—shown in Fig. 3—
will be used where the decoder will be ignored. However, the question is to what extent
the dimensionality can be reduced? Reducing dimensions is tantamount to data loss.
Exaggeration in dimensionality reduction will lead to poor performance.
Weights
Fig. 3. Encoder phase
The exaggeration in reducing numbers of hidden neurons is called under-fitting.

Where the model is incapable of learning the variability of the inputs. Hence, the
penalty of reducing the neurons in hidden layers (dimensionality reduction) is under-
fitting.
The proposed algorithm is reducing the number of the features to a single value,
without suffering the deficiency of under-fitting. The approach adopts full utilization of
the autoencoder network (or RBM), then a single value represents the reconstruction
error between the input and the output is measured.
4 Detection System Architecture
The figure depicts the main components of the proposed detection system. The
framework is established on unsupervised deep neural networks. Two types of USDL
algorithms were used Autoencoders and Restricted Boltzmann machines.
In the context of networks anomalies, threats are continuously growing. Hence,
from a security perspective, an intrusion detection system must be capable of detecting
attacks which have never seen before. This concept is theoretically achievable through
unsupervised learning. Where the supervised algorithms-whether they are classic
algorithms or deep learning must previously see samples and classes to classify the new
records (Fig. 4).
ReconstrucƟ A
on Error
N
K-
Mea N
ns
A
N
Fig. 4. Detection framework
To determine the anomalies using dimensionality reduction techniques, the data

sample is projected on the correlation structure deduced by the algorithm, records with
large—relative to a pre-defined threshold—are marked as anomalies. A similar
approach is theoretically applicable by passing the data through a trained model and
then defines a reconstruction error threshold to isolate anomalies. This is achievable by
a simple regression phase in the second phase. However, our experiments show the
data (reconstruction errors) is non-linear samples. It is not separable by a simple
threshold. Hence, the framework provides a non-linear algorithm to cluster the
reconstruction errors
5 Simulations and Analysis Architecture
The simulation is conducted in two main scenarios based on AE. Figure 5 shows the
main scenario is an implementation for our algorithm based on AE. The sub-scenarios
use two different simple classical algorithms K-means and Shift- means algorithms at
the second phase for clustering. The scenarios occur as following,
• Autoencoder phase then K-means for clustering
• Autoencoder phase then Shift-means for clustering
We used Tensorflow (TF) as a deep learning development library. Tensorflow is
matrices flow in a graph model. TF graph consists of nodes and edges; nodes represent
mathematical operations, edges represent multi-dimensional data arrays (tensors).
Fig. 5. Simulation flowchart
KDD99 is the most used dataset in machine learning and intrusion detection. The
dataset represents real network traffic collected data. The dataset includes 4898431
traffic records for the training, 311029 records for testing [15].
To summarize the results of the simulation, we used confusion matrix statistics. The
confusion matrix consists of columns and rows that list the number of testing samples
as either predicted or actual ratios. Figure 6 is a general description of the confusion
Fig. 6. Confusion matrix

matrix, where we have two classes normal and abnormal. To validate the results the
second phase was done in two different clustering algorithms K-means and shift means.
The confusion matrices for both with different samples sizes are listed Fig. 7. Table 1
shows the AE scored higher than 99% accuracy with K-means
The accuracy = (TP + TN) / (P + N), from Fig. 7 we conclude Table 1 to sum-
marize the accuracy results of each scenario.
Table 1. Accuracy statistics for different scenarios and data samples

800 Sample 400 Sample
Accuracy K-means 0.9988 1
Means-shift 0.9963 0.99
A. AE and K-Means 400 samples B. AE and Shift Means 400 samples
C. AE and K- Means 800 samples D. AE and Shift Means 800 samples
Fig. 7. Confusion matrix statistics

6 Conclusion
Reducing the dimensionality of the data improves the separation of samples into clusters.
However, dimensionality reduction has a limit in autoencoders, where the exaggerations
in reduction negatively affect the model prediction. It causes generalisation deficiencies.
Our approach remarkably reduced the dimension of the data to a single value. We
considered the reconstruction as a representative value. The clustering these values using
a fast algorithm like K-means instead of expensive computational algorithm improves the
performance. Our results show the proposed framework scored accuracy over 99%.
References
1. Ghorbani, A.A., Lu, W., Tavallaee, M.: Network Intrusion Detection and Prevention
Concepts and Techniques. Springer, New York (2010)
2. Mudzingwa, D., Agrawal, R.: A study of methodologies used in intrusion detection and
prevention systems (IDPS). In: 2012 Proceedings of IEEE SoutheastCon, pp. 1–6, 15–18
March 2012
3. Chandola, V., Banerjee, A., Kumar, V.: Anomaly detection for discrete sequences: a survey.
IEEE Trans. Knowl. Data Eng. 24(5), 823–839 (2012)
4. Hinton, G.E., Osindero, S., The, Y.W.: A fast learning algorithm for deep belief nets. Neural
Comput. 18, 1527–1554 (2006). https://doi.org/10.1162/neco.2006.18.7.1527
5. Krizhevsky, A., Sutskever, I., Hinton, G.: ImageNet classification with deep convolutional
neural networks. In: NIPS (2012)
6. Schmidhuber, J.: Deep learning in neural networks: an overview. Neural Netw. 61, 85–117
(2015)
7. Mohamed, A., Dahl, G., Hinton, G.: Acoustic modelling using deep belief networks. IEEE
Trans. Audio Speech Lang. Process. 20(1), 14–22 (2012)
8. Dong, B., Wang, X.: Comparison deep learning method to traditional methods using for
network intrusion detection. In: 2016 8th IEEE International Conference on Communication
Software and Networks (ICCSN), Beijing, pp. 581–585 (2016)
9. Sakurada, M., Yairi, T.: Anomaly detection using autoencoders with nonlinear dimension-
ality reduction. In: Rahman, A., Deng, J., Li, J. (eds.) Proceedings of the MLSDA 2014 2nd
Workshop on Machine Learning for Sensory Data Analysis (MLSDA 2014), p. 4, 8. ACM,
New York (2014). http://dx.doi.org/10.1145/2689746.2689747
10. Fiore, U., Palmieri, F., Castiglione, A., De Santis, A.: Network anomaly detection with the
restricted Boltzmann machine. Neurocomputing 122, 13–23 (2013)
11. Salama, M.A., Eid, H.F., Ramadan, R.A., Darwish, A., Hassanien, A.E.: Hybrid intelligent
intrusion detection scheme. In: Soft Computing in Industrial Applications, pp. 293–303.
Springer, Heidelberg (2011)
12. Zhai, S., Cheng, Y., Lu, W., Zhang, Z.: Deep structured energy based models for anomaly
detection. In: Maria Balcan, F., Weinberger, K.Q. (eds.) Proceedings of the 33rd
International Conference on Machine Learning - Volume 48 (ICML 2016), vol. 48,
pp. 1100–1109 (2016). JMLR.org
13. Tavallaee, M., Bagheri, E., Lu, W., Ghorbani, A.: A detailed analysis of the KDD CUP 99
data set. In: IEEE Symposium on Computational Intelligence for Security and Defense
Applications, CISDA 2009, pp. 1–6 (2009). https://doi.org/10.1109/CISDA.2009.5356528
Blockchain: A Research Framework for Data
Security and Privacy
Farhad Daneshgar, Omid Ameri Sianaki(&),

and Prabhat Guruwacharya
College of Engineering and Science, Victoria University Sydney,

Sydney, Australia
{farhad.daneshgar,omid.amerisianaki}@vu.edu.au,
prabhat.guruwacharya@live.vu.edu.au
Abstract. As an emerging field of research, Blockchain is currently experi-

encing a lack of research frameworks to guide studies in the field. Emerging
technologies are usually triggered first, and then the academic world will attempt
to develop discipline for these technologies. Our current exploratory work-in-
progress is an initial attempt in developing a research framework for investi-
gating the security of data through Blockchain. This study is the first phase of a
more extensive study that aims to provide a conceptual framework for block-
chain research in general. In its current form, the initial proposed framework of
this study can be used for both scoping as well as evaluating existing research
approaches in the field of data security through blockchain technology. It is also
expected to facilitate scoping and categorizing future studies by providing a set
of defined set research categories in the field.
Keywords: Blockchain data security Blockchain research

Blockchain meta data Blockchain research themes
1 Introduction and Research Methodology
Blockchain technology has successfully replaced economic transaction systems in

various organisations, and has the potential to revamp heterogeneous business models
in different industries. Although it promises a securely distributed framework to
facilitate sharing, exchanging, and integration of information across all users and third
parties, it is essential for the planners and decision makers to analyse itblockchain in
depth for its suitability in various industries and business applications [1]. The
blockchain should be deployed only if it is applicable and provides security with better
opportunities for obtaining increased revenue and reductions in cost [2]. The high rate
of blockchain popularity is rooted in its unique characteristics including decentralized
infrastructure and peer-to-peer nature.
Currently, many blockchain categories of research is being overshadowed by
Bitcoin, whereas blockchain could be applied to a variety of fields far beyond Bitcoin;
and the current study is one initial attempt to further explicate both the research and
application landscape of this exciting technology. One such emerging field is the data
security and privacy that is the subject of the current study. Blockchain has numerous

https://doi.org/10.1007/978-3-030-15035-8_95
Blockchain: A Research Framework for Data Security and Privacy 967
benefits such as decentralisation, persistency, anonymity, and auditability. Blockchain

applications range from cryptocurrency, financial services, risk management, internet
of things (IoT), and various services [3]. Because of these unique characteristics, the
blockchain technology is becoming an important technology for the next generation of
internet interaction systems particularly smart contracts [4], public services [5], Internet
of Things (IoT) [6], reputation systems [7], and security services [8]. In the current
work-in-progress research, we attempt to identify various themes/dimensions that
collectively define the research field of data security and privacy through blockchain.
More specifically, we are interested in understanding how past and present researchers
have been investigating the above topic of interest, and what approaches the latter
adopt in their research. We aim to explore what are the relevant contextual knowledge
domains or meta data that can guide current and future researchers through their works.
In our study, each such dimension is called a ‘theme’, as explained in the next section.
With the topic ‘data security through blockchain’ being an emerging research field
with an unavoidable less rigorous academic discipline, one can expect that the previous
studies may have been based, more or less, on particular theme decided upon by the
front-runner researchers. As an example, a particular study may overlap two areas of
blockchain algorithm and blockchain data security. At this early stage of the blockchain
hype cycle (speaking in Gartner’s tongue), it may not be clear as to what category, meta
data, or keyword would best describe that particular research. The current study pro-
vides an overarching theme called ‘blockchain data security and privacy’, and within
that main category, various sub-categories such as ‘blockchain (security) applications’,
‘blockchain (security) algorithm’, and other relevant categories are fitted.
In the current study, we have explored various fundamental themes and present
them as a set of meta data or ‘themes’ that can potentially provide a frame of work for
classifying current and future studies in the field of blockchain security. The innovation
component of our research lies in the categorisation scheme and explored themes that
have been synthesized from the current literature. The research methodology adopted
for the current study is conceptual systematic review, and is explained later on in more
details. The proposed research framework is expected to further explicate the current
‘understanding’ of the study research topic including its scope and dimension, that in
turn, will facilitate in what areas the future understandings can be suggested. More
specifically, it is expected that the proposed research framework be used to (i) provide
the researchers with a frame to both scope and present their works, and (ii) encourage
and facilitate future studies by providing a set of themes representing bigger picture
surrounding the topic of data security through blockchain. And finally, we acknowl-
edge that our current study is in line with many recent efforts by several government
agencies that are in the process of developing a roadmap for blockchain standards.
A recent document published by Standards Australia highlights the critical role that
Australia will play in leading international efforts to develop blockchain standards
under ISO/TC 307 Blockchain and electronic distributed ledger technologies [9].
The remaining parts of the paper are structured as follows: The next section is
allocated to the literature review with two groups of reviews, one related to the tech-
nological and application perspectives of the blockchain in general. This review
identifies the scope of the current work-in-progress within the overall blockchain
landscape research. The second group of reviews constitutes the bulk of reviews for the
968 F. Daneshgar et al.
current and future studies, and is titled ‘Themes of Data Security and Privacy through
Blockchain’. This part of the review will focus on synthesising various existing
blockchain studies that directly relate to the use of blockchain technology for the
maintaining data security and privacy. The Literature Review section is followed by the
last section that provides concluding remarks along with highlight of the authors’ future
work.
2 Literature Review
As stated before, this study identifies various themes and dimensions that collectively
define and scope current and future research in data security through blockchain.
Initially, the proposed research framework is based on the existing body of literature in
the fields, and the themes that the front-runner researchers have implied for their
research. For the current study, we have explored a set of initial themes as a starting
point. As one of the authors’ future work, the proposed research framework of the
current study will be updated cumulatively, as additional works are found in the
literature; hence the concept dynamic meta thematic model for blockchain research. To
achieve the above goal, we explored the relevant current knowledge landscape in order
to search for existing research works by early researchers that either directly or indi-
rectly relate to the topic of data security through blockchain. The current study partly
achieves this goal. More specifically, in an attempt to construct our proposed research
framework, we conduct a two-round review of the literature. In the first round, that is
the current study; we develop an initial categorisation scheme by identifying possible
themes or categories that define the current body of the literature in the field. In the
second round of review, in our future study, each existing study will be assessed by the
proposed framework to either be matched with an existing category/theme, or create a
new category/them, or update an existing one. The authors acknowledge the dynamic
nature of the proposed research framework, and for this reason, they intend to use the
current study as a basis for attracting collaborating researchers and future research
grants for maintaining the currency of the proposed research framework for all aspects
of Blockchain research.
The research methodology adopted for the current study is a conceptual systematic
review where the search criteria is a specific concept, and the goal of the review is to
develop a conceptual framework to guide empirical data collection and analysis [10].
To apply this methodology for the current study, we have collected relevant research
articles and major industry reports, and grouped them according to a set of pre-defined
concepts called categories or themes. In the current study, the primary theme at the
highest level is blockchain-based security, that has been used in the first round of the
review. During this first round of the review, we identified two sets of relevant research
categories: (i) blockchain technological and application perspectives, and (ii) specific
themes related to the data security and privacy through blockchain. These two are
discussed below:
3 Technological and Application Perspective
As an introduction to our proposed research framework, we provide this section to

scope our own work at a higher-level perspective. One comprehensive study provides
an overarching framework for the blockchain research as a whole while adopting a dual
technological-application perspective [3]. The above study provides the blockchain
taxonomy, blockchain consensus algorithms, blockchain applications, technical chal-
lenges, and recent advances in tackling the challenges. The above study also identifies
five major blockchain application categories including finance, Internet of Things (IoT),
public and social services, reputation system, and security and privacy [3]. Based on
the above scheme, the subject matter of the current study will fit into the blockchain
security and privacy application. The above study also identifies three typical chal-
lenges and problems facing the blockchain technology including scalability, privacy
leakage, and selfish mining. Discussions on these themes fall outside the scope of the
current study. The next sections provide more focused review of the literature with the
aim of identifying various themes of the framework.
4 Themes of Data Security and Privacy Through Blockchain
Theme 1: Blockchain as a Distributed Database. This theme focusses on the dis-

tributed nature of the blockchain as an enabler of data security and privacy. Under this
approach, blockchain is considered as the public ledger for transactions, and it prevents
hacking during transactions involving virtual cash. As a type of distributed database
and a data record list that continuously grows, it is designed to disable arbitrary
tampering by the operator of distributed peers [11]. Transaction records are encrypted
according to a rule and operated in computers that run the blockchain software. As a
result, and due to its secure data distribution capability, blockchain has become a
foundation for the next-generation financial technology [11]. Blockchain provides
security through the authentication of peers that share virtual cash, encryption, and the
generation of hash value. There seem to be an overlap between this theme and theme 5,
security algorithms of the blockchain, and future updates to the proposed research
framework is expected to resolve the issue either by merging existing themes, or by
creating new themes.
Theme 2: Blockchain as a Technology to Secure Distributed Networks. Currently,
there is a variety of anti-malware filters for detection of the suspected files through
pattern-matching schemes. These pattern-matching schemes themselves are based on a
central server to store and update the virus patterns. However, these centralised
counter-measures are also vulnerable to malicious attackers [3]. Blockchain can
potentially help to improve the security of distributed networks. A recent study [10]
proposed a novel anti-malware environment named BitAV, in which users can dis-
tribute the virus patterns on blockchain in order for the BitAV to enhance the fault
tolerance. BitAV can improve the scanning speed and enhance the fault reliability (i.e.,
less susceptible to targeted denial-of-service attacks). Furthermore, according to Axon
[12], blockchain can be used to construct a privacy-aware PKI (Public Key
Infrastructures) while simultaneously improving the reliability of conventional PKIs.

Zyskind et al. [13] proposed a blockchain-based decentralized personal data manage-
ment system that maintains user ownership of the data.
Theme 3: Blockchain (BC)-Based Architecture for Preserving Data Privacy. The
blockchain is a sequence of blocks, which holds a complete list of transaction records
like conventional public ledger [14]. Each block points to its immediately previous
block via a reference that is essentially a hash value of the previous block called parent
block. The first block of a blockchain is called genesis block which has no parent block
[15]. A block consists of the block header and the block body. The header includes
block version, parent block hash, Merkle tree root hash, timestamp, nBits, and Nonce.
Further details on the architectural aspects of the blockchain can be found at [14] and
[15]. The theme BC-based architecture focusses on the architecture design for main-
taining data security through Blockchain. One such blockchain-based architecture is
utilized for smart vehicles. Inter-connected services in smart vehicles offer sophisti-
cated benefits to all stakeholders involved; however, these services expose the smart
vehicles and their users to a range of security and privacy threats such as location
tracking or remote hijacking of the vehicle. A study by Dorri et al. [16] proposes BC-
based architecture that protects the privacy of the users of smart vehicles, and to
increase the security of the vehicular ecosystem. Wireless remote software updates
other emerging and relevant services such as dynamic vehicle insurance fees. By
relying on the generic characteristics of the blockchain technology as outlined before,
the above authors qualitatively argue the resilience of their proposed BC-based
architecture against common security attacks.
Despite all the above blockchain benefits as a result of its inherent distributed nature,
the blockchain has been criticized for imposing potential risks on the identity of the
users. In a recent study, Primavera De Filippi [17] argues that on the one hand, the
pseudonymous nature of many blockchain-based networks allows for people to transact
on a peer-to-peer basis, without disclosing their identity to anyone. On the other hand,
the transparency inherent to these networks is such that anyone can retrieve the history
of all transactions performed on a blockchain and rely on big data analytics in order to
retrieve potentially sensitive information. As a result, in order to understand the privacy
implications of the blockchain’s decentralised architecture, we must analyse whether
the privacy gains resulting from decentralised coordination are more significant than
the privacy costs derived from the disclosure of metadata that may reveal personal
information [Ibid]. Another very recent study provides two cases related to the identity
of shoppers. The study mentions that on most shopping websites, third-party trackers
receive information about user purchases for purposes of advertising and analytics.
However, if the user pays using a cryptocurrency, trackers typically possess enough
information about the purchase to identify the transaction on the blockchain uniquely
and link it to the user’s cookie, and further to the user’s real identity [2].
Theme 4: Data Security Applications of Blockchain. Cloud data provenance is
metadata that records the history of the creation and operations performed on a cloud
data object; and as a result, secure data provenance is crucial for data accountability,
forensics and privacy [18]. A recent article proposes a decentralised and trusted cloud
data provenance using blockchain technology [19]. The above researchers demonstrate
that blockchain technology improves the performance of the blockchain-based data

provenance by providing tamper-proof records, maintaining data transparency and
accountability in the cloud, and helps to enhance the privacy and availability of the
provenance data [Ibid]. The above researchers designed and implemented ProvChain,
an architecture to collect and verify cloud data provenance, all by embedding the
provenance data into blockchain transactions. ProvChain operates mainly in three
phases: (1) provenance data collection, (2) provenance data storage, and (3) provenance
data validation. The above authors further demonstrate that ProvChain provides
security features including tamper-proof provenance, user privacy and reliability with
low overhead for the cloud storage applications [19].
Another exciting application is the use of blockchain technology for the creation of
smart homes [20]. These homes are equipped with always online device ‘miner’
responsible of handling all the communications. The study demonstrates that the smart
home framework is secure by methodically examining its data security with regards to
the fundamental security goals of CIA [20].
As yet another application, one study demonstrates how blockchain capabilities can
be exploited to create decentralized, shared economy applications such as Airbnb and
Uber “that allow people to monetise, securely, their things to create more wealth.”
[21]. The above authors argue that with the recent interest in the Internet of Things and
blockchain, the opportunity exists for creating a myriad of sharing applications such as
peer-to-peer automatic payment mechanisms, foreign exchange platforms, digital rights
management, and cultural heritage, to name but a few [21].
Theme 5: Security Algorithms of Blockchain. In blockchain, how to reach con-
sensus among the untrustworthy nodes is a transformation of the Byzantine Generals
(BG) Problem (see [22, 23] for more details on BG). In blockchain, no central node
ensures ledgers on distributed nodes which are all the same. Nodes need not to trust
other nodes. Thus, some protocols are needed to ensure that ledgers in different nodes
are consistent. There are several common approaches to reach consensus in the
blockchain [3]. These are briefly explained below:
Proof of work (PoW) is a consensus strategy used in the Bitcoin network [23].
POW requires a complicated computational process in the authentication. In POW,
each node of the network calculates a hash value of the continually changing block
header. The consensus requires that the calculated value must be equal to or smaller
than a particular given value.
Proof of stake (PoS) is an energy-saving alternative to POW. Instead of demanding
users to find a nonce in unlimited space, POS requires people to prove the ownership of
the amount of currency because it is believed that people with more currencies would
be less likely to attack the network [22]. However, the selection based on account
balance is quite unfair because the single richest person is bound to be dominating the
network. As a result, many solutions are proposed with the combination of the stake
size to decide which one to forge the next block [Ibid].
Practical Byzantine Fault Tolerance (PBFT) is a replication algorithm to tolerate
Byzantine faults [22]. Hyperledger Fabric [24] utilises the PBFT as its consensus
algorithm since PBFT could handle up to 1/3 malicious byzantine replicas. A new
block is determined in one round. In each round, a primary would be selected according
to some rules. Moreover, it is responsible for ordering the transaction. The whole
process could be divided into three phases: pre-prepared, prepared, and commit. In
each phase, a node would enter next phase if it has received votes from over 2/3 of all
nodes. So, PBFT requires that every node is known to the network.
Stellar Consensus Protocol (SCP) [25] is also a Byzantine agreement protocol.
There is no hashing procedure in PBFT. In PBFT, each node has to query other nodes
while SCP gives participants the right to choose which set of other participants to
believe. Based on PBFT, Antshares [26] has implemented their delegated byzantine
fault tolerance (dBFT). In dBFT, some professional nodes are voted to record the
transactions instead of all nodes.
Delegated proof of stake (DPOS) is a method similar to POS. In this algorithm,
miners get their priority to generate the blocks according to their stake. The significant
difference between POS and DPOS is that POS is a direct democratic while DPOS is a
representative democracy. Stakeholders elect their delegates to generate and validate a
block. With significantly fewer nodes to validate the block, the block could be con-
firmed quickly, making the transactions confirmed quickly. Meanwhile, the parameters
of the network such as block size and block intervals could be tuned. Additionally,
users do not need to worry about the dishonest delegates because the delegates could be
voted out quickly. Ripple is a consensus algorithm that utilises collectively-trusted
subnetworks within the more extensive network [27]. In the network, nodes are divided
into two types: a server for participating consensus process and client for only trans-
ferring funds. In contrast to that PBFT nodes have to ask every node in the network,
each Ripple server has a Unique Node List (UNL) to query.
5 Concluding Remarks and Future Work
In this exploratory study, we developed an initial version of a research framework that

can be used for researchers in the field of security of data through Blockchain as an
analytical tool for scoping and categorising their work within the blockchain security
and privacy research landscape. The focus of the proposed framework has been to
represent the current research landscape in the field by various meta data in the form of
classified themes. The ultimate goal of the current study is to use this initial set of
categories as precursor to a more extensive future study that, among other things, can
be used to facilitate the distinction between the hype cycle and the maturity of the
blockchain technology. The current work-in-progress is the first phase of a more
extensive future study that aims to provide a dynamic conceptual framework for
describing the phenomena over time. The proposed framework is expected to be used
for both scoping as well as evaluating existing studies in the field. The proposed
framework is also expected to facilitate future studies by providing a defined academic
framework for future researchers in the field.
The study identified five overlapping themes including blockchain as a distributed
database, blockchain as a technology to secure distributed networks, a blockchain-
based architecture for preserving data privacy, data security applications of the
blockchain, and security algorithms of the blockchain.
In the next phase of our study, we plan to use the proposed research framework of
the current study and periodically reassess all upcoming works in the field of data
security through blockchain, and as a result, to enhance the proposed framework. Such
enhancement is defined by a need to either updating some of the existing themes or by
adding new themes to the research landscape. These iterative enhancements will
continue until the point where the last group of studies do not bring a major change to
the research framework. At that stage, one can claim to settle down of the emerging
blockchain technology about the security issues at academic levels.
References
1. Puthal, D., Malik, N., Mohanty, S.P., Kougianos, E., Yang, C.: The blockchain as a
decentralized security framework. IEEE Consum. Electron. Mag. 7, 18–21 (2018)
2. Goldfeder, S., Kalodner, H., Reisman, D., Narayanan, A.: When the cookie meets the
blockchain: privacy risks of web payments via cryptocurrencies. Proc. Priv. Enhanc.
Technol. 2018, 179–199 (2018)
3. Zheng, Z., Xie, S., Dai, H.-N., Chen, X., Wang, H.: Blockchain challenges and
opportunities: a survey. Int. J. Web Grid Serv. 14, 352–375 (2018)
4. Kosba, A., Miller, A., Shi, E., Wen, Z., Papamanthou, C.: Hawk: the blockchain model of
cryptography and privacy-preserving smart contracts. In: 2016 IEEE Symposium on Security
and Privacy (SP), pp. 839–858 (2016)
5. Akins, B.W., Chapman, J.L., Gordon, J.M.: A whole new world: income tax considerations
of the Bitcoin economy. Pitt. Tax Rev. 12, 25 (2014)
6. Zhang, Y., Wen, J.: An IoT electric business model based on the protocol of bitcoin. In:
2015 18th International Conference on Intelligence in Next Generation Networks (ICIN),
pp. 184–191 (2015)
7. Sharples, M., Domingue, J.: The blockchain and kudos: a distributed system for educational
record, reputation and reward. In: European Conference on Technology Enhanced Learning,
pp. 490–496 (2016)
8. Noyes, C.: Bitav: fast anti-malware by distributed blockchain consensus and feedforward
scanning. arXiv preprint arXiv:1601.01405 (2016)
9. Standards Australia: Roadmap for blockchain standards report (2017)
10. Seuring, S., Müller, M.: From a literature review to a conceptual framework for sustainable
supply chain management. J. Clean. Prod. 16, 1699–1710 (2008)
11. Park, J.H., Park, J.H.: Blockchain security in cloud computing: use cases, challenges, and
solutions. Symmetry 9, 164 (2017)
12. Axon, L.: Privacy-awareness in blockchain-based PKI (2015)
13. Zyskind, G., Nathan, O.: Decentralizing privacy: using blockchain to protect personal data.
In: 2015 IEEE Security and Privacy Workshops (SPW), pp. 180-184 (2015)
14. Chuen, D.L.K.: Handbook of Digital Currency: Bitcoin, Innovation, Financial Instruments,
and Big Data. Academic Press, Cambridge (2015)
15. Zheng, Z., Xie, S., Dai, H., Chen, X., Wang, H.: An overview of blockchain technology:
architecture, consensus, and future trends. In: 2017 IEEE International Congress on Big Data
(BigData Congress), pp. 557-564 (2017)
16. Dorri, A., Steger, M., Kanhere, S.S., Jurdak, R.: Blockchain: a distributed solution to
automotive security and privacy. IEEE Commun. Mag. 55, 119–125 (2017)
17. De Filippi, P.: The interplay between decentralization and privacy: the case of blockchain
technologies (2016)
18. Sarwar, M.U., Hanif, M.K., Talib, R., Sarwar, B., Hussain, W.: Data provenance for cloud
computing using watermark. Int. J. Adv. Comput. Sci. Appl. 8, 407–411 (2017)
19. Liang, X., Shetty, S., Tosh, D., Kamhoua, C., Kwiat, K., Njilla, L.: Provchain: a blockchain-
based data provenance architecture in cloud environment with enhanced privacy and
availability. In: Proceedings of the 17th IEEE/ACM International Symposium on Cluster,
Cloud and Grid Computing, pp. 468–477 (2017)
20. Dorri, A., Kanhere, S.S., Jurdak, R., Gauravaram, P.: Blockchain for IoT security and
privacy: the case study of a smart home. In: 2017 IEEE International Conference on
Pervasive Computing and Communications Workshops (PerCom Workshops), pp. 618–623
(2017)
21. Huckle, S., Bhattacharya, R., White, M., Beloff, N.: Internet of things, blockchain and shared
economy applications. Proc. Comput. Sci. 98, 461–466 (2016)
22. Castro, M., Liskov, B.: Practical Byzantine fault tolerance. In: OSDI, pp. 173–186 (1999)
23. Nakamoto, S.: The proof-of-work chain is a solution to the Byzantine generals’ problem.
The Cryptography Mailing List, November 2008
24. Cachin, C.: Architecture of the hyperledger blockchain fabric. In: Workshop on Distributed
Cryptocurrencies and Consensus Ledgers (2016)
25. Mazieres, D.: The stellar consensus protocol: a federated model for internet-level consensus.
Stellar Development Foundation (2015)
26. Zhu, H., Zhou, Z.Z.: Analysis and outlook of applications of blockchain technology to
equity crowdfunding in China. Financ. Innov. 2, 29 (2016)
27. Schwartz, D., Youngs, N., Britto, A.: The ripple protocol consensus algorithm. Ripple Labs
Inc White Paper (2014)
Environmental Monitoring Intelligent System
Using Wireless Nanosensor Networks
Nedal Ababneh(&), Omid Ameri Sianaki, and Shafquat Hussain
College of Engineering and Science, Victoria University, Sydney, Australia

{Nedal.Ababneh,Omid.AmeriSianaki,
Shafquat.Hussain}@vu.edu.au
Abstract. Wireless nanosensor networks are envisioned to operate in the THz

band, due to the small size of the network devices. The tiny dimensions of the
nanosensor devices allow them to be embedded in various objects and devices in
our environment to obtain fine-grained data, such as object’s components or its
internal status, from hard-to-reach locations in a non-invasive way, and to
perform additional in-network processing and thereby enabling a myriad of
novel applications in industrial, biomedical, and agricultural settings that cannot
be realized with conventional sensor networks. In this work, we propose a crop
monitoring/defense application of nanosensors, for detection of any compound
released by plants. Based on the detected volatile compounds, the state of the
plant in its surrounding can be recognized and chemical nanosensors can release
the same natural composites to reinforce the plants defense mechanism in case
of insect attack. Our approach centers around the creation of a wireless
nanosensor network (WNSN) wherein large number of chemical nanosensor
devices, which is connected to a remote center via the Internet to provide remote
monitoring/controlling capabilities.
1 Introduction
Advances in nanotechnology have enabled the development of sensor devices in a scale

ranging from one to a few hundred nanometers. At this scale, a nanosensor is defined as
the most basic functional unit, integrated by nano-components and able to perform
simple tasks such as computing, sensing, actuation, and communication [1]. In contrast
to traditional Wireless Sensor Networks [2–4] and Wireless Body Area Sensor Net-
works [5, 6], The tiny dimensions of the nanosensor devices allow them to be
embedded in various objects and devices in our environment to obtain fine-grained
data, such as object’s components or its internal status, from hard-to-reach locations in
a non-invasive way, and to perform additional in-network processing and thereby
enabling a myriad of novel applications in industrial, biomedical, and agricultural
settings that cannot be realized with conventional sensor networks. Although
nanosensors are not yet available commercially, there have been significant relevant
developments in recent years that point to a future when such devices could be pro-
duced in bulk. Yonzon et al. [7] surveyed many types of nanosensors that can be used
for chemical and biological sensing. Also, progress has been recorded in chemical and
biological nanoactuators that can be used to accomplish some basic tasks at the

https://doi.org/10.1007/978-3-030-15035-8_96
976 N. Ababneh et al.
molecular level by harnessing interactions between nanoparticles, electromagnetic

fields, and heat [8]. Researchers envisaged nanosensors fabricated from Carbon nan-
otubes that can act as highly accurate chemical sensors [9]. Moreover, a graphene-
based nanoantenna can be integrated into the structure of the nanosensor device; such
antennas radiate in an extremely high frequency band in the THz region, 0.3–10 THz
[10]. The theoretical bandwidth available to communication based on graphene can be
quite high; however, many constraints emerge within this band that are not present at
lower frequencies. The most striking characteristic of the Terahertz channel is that of
molecular absorption of electromagnetic waves (the terahertz is also the resonant fre-
quency of many molecules) and the subsequent generation of noise.
In this project, PlantMONT, we use nanosensors devices for high-resolution agri-
cultural crop monitoring/defense. It is known that plants release specific composites in
the air at their blooming stage, they also emit an extensive set of chemical compounds
above and below the ground, providing a means of communication through which they
share vital information about their surrounding environment, e.g., available resources,
physical damages or insect attacks. In the case of detecting an insect attack, some plant
species emit various volatile organic compounds with different concentrations, based
on the insect type. These emissions carry information on the attack of the plant and
seek to attract the natural predators of the attacking insects as well as activating/
strengthening other plants’ defense systems [11, 12].
Fig. 1. Network hierarchy of a plant monitoring nanosensor network, composed of chemical

nanosensors, macrosinks (routers), and a gateway.
Environmental Monitoring Intelligent System 977
In PlantMONT, we propose a crop monitoring/defense application of nanosensors,

for detection of any compound released by plants. Based on the detected volatile
compounds, the state of the plant in its surrounding can be recognized and chemical
nanosensors can release the same natural composites to reinforce the plants’ defense
mechanism in case of insect attack. Our approach centers around the creation of a
wireless nanosensor network (WNSN) wherein large number of chemical nanosensor
devices; each equipped with a chemical nanosensor unit that can detect the presence
and concentration of specific molecules in the environment with high accuracy,
deployed on plant leaves communicate the sensed data to macrosinks (routers) with a
attached to the plant stem. The aggregated data will be forwarded to a gateway device,
which is connected to a remote center via the Internet to provide remote
monitoring/controlling capabilities. The proposed system architecture is depicted in
Fig. 1. By considering a hierarchical structure for WNSNs, their management will be
easier compared to an ad-hoc structure e.g., routing, addressing, or network discovery.
Additionally, such a topology provides the possibility of connecting nanosensor net-
works to other external networks. The nanosensor network is composed of Nanode-
vices (nanosensors), Macro/Micro-devices (macrosinks/nanorouters) and Macroscale
gateways (nanointerfaces) [13, 14].
The remainder of this paper is organized as follows. Section 2 presents the main
objectives and contributions of PlanMONT project. Section 3 provides the detailed
implementation and work plan. Finally, Sect. 4 concludes the paper.
2 PlantMONT
The expected capabilities and peculiarities of NSNs and the terahertz channel require
the development of novel, reliable and yet energy efficient medium access and com-
munication protocols. In this project, new medium access control (MAC), addressing,
network discovery and association and routing protocols will be designed, imple-
mented and investigated. PlantMONT has the following concrete objectives:
2.1 Channel Sharing

MAC protocols used in traditional networks, usually based on carrier sensing (e.g.,
CSMA and all its variations) need to be revised; they cannot be used in pulse-based
communications because there is no carrier signal to sense. Within this project, we will
propose, design and investigate novel MAC protocols specifically suited for NSNs
taking into account the limited energy and time-varying communication channel of
such networks. In NSNs, the transmitted packets are very short, and the probability of
collision between nanosensor nodes trying to access the channel at the same time will
be much lower than in traditional wireless sensor networks. In some scenarios, it is
even possible that the probability of collision may be neglected, leading to the inte-
gration of MAC protocols in the physical layer. Having said this, we will design and
investigate novel contention-based (random access) and scheduling-based (channel
partitioning/time-division multiple access) MAC protocols. We will formulate the
scheduling problem as a mathematical optimization problem and propose a Mixed
Integer Linear Program (MILP) model to solve it. There are few attempts in the
literature to address multi-access and channel sharing problem in NSNs environment,
however most of the research do not take into consideration the unique characteristics
of such networks and thus, need to be improved further. Jornet et al. suggested the first
MAC protocol for WNSN, called PHLAME [15]. The proposed framework design is
based on a new modulation schema called RD TS-OOK, an extended version of TS-
OOK [16]. PHLAME uses the advantages of the low-weight channel coding protocol
proposed in [17]. By reducing the code weight, that is, encoding the information using
more logical 0’s than logical 1’s, both molecular absorption noise and interference can
be mitigated without affecting the throughput of the WNSNs. PHLAME provides joint
selection of the optimal communication parameters and channel coding scheme for
both the sender and the receiver. However, implementation feasibility and energy
efficiency evaluation of the method are still open questions: First, the network scala-
bility might be an issue for PHLAME, as performance of the proposed approach has
only been investigated for a maximum of 10 nanomotes per square millimeter, which
might not be a realistic scenario for some applications where thousands of nanosensors
per square millimetre are needed. Second, PHLAME relies on the low-weight channel
coding, which might not be applicable for WNSNs in some environments where the
codeword is hard to define for the transmitted data due to natural randomness in the
environment. Finally, the handshake process in the PHLAME may exhaust the energy
of nanosensors, limiting the throughput of the network, and nanosensors might not
have enough computational resources to dynamically find the optimal communication
parameters. To alleviate the last restriction of the PHLAME, Wang et al. extended the
original PHLAME to increase the throughput and lifetime of WNSNs by jointly
optimizing the energy harvesting and consumption processes [18]. In other research,
Piro et al. attempted to adopt the existing MAC protocols, mainly ALOHA and carrier
sense multiple access (CSMA), for WNSNs and investigate their performances [19].
2.2 Addressing, Network Association and Service Discovery

In NSNs, assigning a unique address to every nanosensor (and nanorouter) is not a
simple task, mainly due to the fact that this would require complex synchronization and
coordination between nanosensors. Moreover, taking into account that every single
NSN will already contain thousands of nanosensors, the inter-networking of all of them
would require the use of very long addresses. However, some simpler and more fea-
sible alternatives need to be developed. For example, taking into account the hierar-
chical network architecture shown in Fig. 1, we can relax the previous requirement and
only force those nanosensors coordinated by the same nanorouter to have different
addresses. More interestingly, we can think of specific applications in which it is not
necessary to have information from a specific nanosensor, but, for example, from a type
of nanosensors (data-centric approach). In particular, different type of components may
have different addresses, but identical nanosensors can behave in the same way in terms
of communication; different nodes will react in the same way depending on the
information that is being sensed or their internal state. This can relax the coordination
requirements among nanosensors, while still supporting interesting applications. In
addition, these same concepts can be used to develop new network discovery and
network association protocols for nanosensors, where every nanosensor is expected to

be able to seamlessly connect to the network and at the same time inform the other
devices about its presence. Taking into account the number of nanosensors that can be
involved in such a network, new network association and service discovery solutions
are needed and will be developed in this project. In our application, it will not be
necessary to notify the entire network when a new nanosensor joins, but just the closer
nanorouter or macro-interface at most.
2.3 Routing
Nanosensors may have to answer a specific query from a command center or may need
to report new events in a push-based fashion. This flow of information (in two-
directions) requires the establishment of routes. Due to the very limited transmission
range (in the order of few tens of millimeters) of nanosensors, multi-hop communication
will be the standard way to communicate between senders and receivers. It is worth to
note that the creation and the management of multi-hop paths from each nanosensor to
its reference nanorouter (cluster head or macrosink) is a very challenging open issue,
which could be very critical in those scenarios entailing some degree of mobility. Tiny
nanosensor devices randomly deployed in an open field can be easily taken away by an
air flow. This may be seen as a shortcoming but can also be exploited in a beneficial
way. All main aspects related to the communications among nanosensors, the interac-
tions between them and nanorouters, and the interface/gateway between a WNSN and
the Internet (command center) will be deeply studied in this project to discover the most
promising and feasible solutions. We will investigate novel Inter- and Intra-network
(plant) routing mechanisms that consider the particular energy requirements and
mobility pattern of nansensors. Also, End-to-end reliability in NSNs has to be guar-
anteed both for the packets going from a remote command center to the nanosensors, as
well as for the packets coming from the nanosensors to a common nanosink (nanor-
outer). We will investigate multi-path routing, selecting the optimal path, forming
clusters, nanorouter/sink optimal placement, and data dissemination and aggregation
techniques. We will design a multi-objective function to address several routing metrics
such as energy consumption, end-to-end delay, throughput, and communication relia-
bility. Pierobon et al. proposed the first routing protocol for WNSNs to optimize the use
of harvested energy to guarantee continuous operation of WNSNs while improving the
overall network throughput [20]. Due to the resource restriction of the nanosensors, in
this work, a hierarchical cluster-based structure is used in which the WNSN is formed
from a few clusters, and each cluster is coordinated by a nanoscale controller, a nano
device with more advanced capabilities than a nanosensor. Later, Liaskos et al. proposed
a metaheuristic-based, selective flooding dissemination scheme for nano-networks [21],
which was later refined in terms of complexity [22]. Tsioliaridou et. al. proposed a joint
peer-to-peer routing and geo-addressing approach, CORONA, for NSNs [23]. COR-
ONA is a stateless and scalable nano-routing approach. It proposes a packet flood
approach, contained within the area defined by the minimal and maximal anchor dis-
tances of the communicating node-pair. However, routing protocol design for nano-
networks is in its very early stages and lots of effort needed here to address the different
routing issues. As stated above, nanosensors are not commercially available yet, and
thus, real experimentations on nanosensor devices are not possible. Moreover, since
WNSNs are too complex for traditional analytical methods to provide an accurate
modelling and understanding of system behavior, network simulators (i.e., ns-2 and
Nano-Sim based on ns-3) and Mathematical Programming-based optimization tech-
niques such as ILP and constraint programming will be used in this project. The ns-2 and
ns-3 simulators are the most prominent tools to develop network, protocol and topology;
they provide substantial supports for designing physical layer, MAC, and Routing
protocols, and investigating how various protocols perform with different configurations
and topologies in a controlled manner. To the best of our knowledge, these two
methodologies/approaches are never been used in the literature so far.
3 Implementation
Three Work Packages (WP) are necessary to achieve the research goals of PlantMONT
(detailed in Sect. 2), and dedicate one other to dissemination and exploitation (and
public engagement) of results:
WP1: This work package deals with the development of the models (Objective 2.1).
WP2: This work package deals with the development of models and solutions for
objectives 2.2 and partially 2.3.
WP3: Having a good understanding of MAC, intra-network routing/dissemination,
and network topology/deployment, we are ready to look at (Objective 2.3):
4 Conclusion
Wireless nanonetworks are a new generation of networks at nano-scale, which pro-

mises new solutions for many applications in the healthcare and environmental mon-
itoring fields. There is a tremendous number of works addressing the physical layer and
some on the link layer. However, medium access and routing in nanosensor networks
have been partially investigated and a lot of effort has to be done here before it becomes
a reality. In this project, we present an environmental monitoring system using
nanosensor network operates at THz frequency band. In particular, we investigate,
design and develop new medium access control (MAC), optimal relaying, addressing
and routing protocols, which captures the peculiarities of the THz channel in order to
build a more reliable, controllable and energy efficient monitoring applications.
References
1. Akyildiz, I.F., Miquel Jornet, J.: Electromagnetic wireless nanosensor networks. Nano
Commun. Netw. 1(1), 3–19 (2010)
2. Ababneh, N., Viglas, A., Labiod, H., Boukhatem, N.: ECTC: energy efficient topology
control algorithm for wireless sensor networks. In: Proceedings of the IEEE International
Symposium on a World of Wireless Mobile and Multimedia Networks (WoWMoM 2009),
Kos Island, Greece, June 2009
3. Ababneh, N.: Radio irregularity problem in wireless sensor networks: a new experimental
results. In: Proceedings of the 32nd IEEE Sarnoff Symposium (Sarnoff 2009), Princeton,
USA, April 2009
4. Ababneh, N., Ortiz, A.M., Timmons, N., Morrison, J.: Energy-efficient multicast tree
construction protocol for real-time data streaming in WSNs. In: Proceedings of the IEEE
International Conference on Wireless and Mobile Computing, Networking and Communi-
cations (WiMob 2012), Barcelona, Spain, 8–10 October 2012
5. Ababneh, N., Timmons, N., Morrison, J.: An adaptive bandwidth allocation scheme for data
streaming over body area networks. In: Proceedings of WWIC (2012)
6. Ababneh, N., Timmons, N., Morrison, J.: Cross-layer optimization protocol for guaranteed
data streaming over wireless body area networks. In: Proceedings of the International
Wireless Communications and Mobile Computing Conference (IWCMC 2012), Cyprus, 27–
31 August 2012
7. Yonzon, C.R., Stuart, D.A., Zhang, X., McFarland, A.D., Haynes, C.L., Van Duyne, R.P.:
Towards advanced chemical and biological nanosensors—an overview. Talanta 67(3), 438–
448 (2005). Nanoscience and Nanotechnology
8. Li, C., Thostenson, E.T., Chou, T.W.: Sensors and actuators based on carbon nanotubes and
their composites: a review. Compos. Sci. Technol. 68(6), 1227–1249 (2008)
9. Cui, Y., Wei, Q., Park, H., Lieber, C.M.: Nanowire nanosensors for highly sensitive and
selective detection of biological and chemical species. Science 293(5533), 1289–1292
(2001)
10. Jornet, J., Akyildiz, I.F.: Graphene-based nano-antennas for electromagnetic nanocommu-
nications in the terahertz band. In: European Conference on Antennas and Propagation
(EuCAP), pp. 1–5 (2010)
11. Heil, M., Bueno, J.C.S.: Within-plant signaling by volatiles leads to induction and priming
of an indirect plant defense in nature. Proc. Natl. Acad. Sci. U.S.A. 104, 5467–5472 (2007)
12. Heil, M., Ton, J.: Long-distance signalling in plant defence. Trends Plant Sci. 13(6), 264–
272 (2008)
13. Afsharinejad, A., Davy, A., Jennings, B.: Dynamic channel allocation in electromagnetic
nanonetworks for high resolution monitoring of plants. Nano Commun. Netw. 7, 2–16
(2016)
14. Akyildiz, I.F., Jornet, J.M.: The Internet of nano-things. IEEE Wirel. Commun. 17(6), 58–63
(2010)
15. Jornet, J.M., Pujol, J.C., Pareta, J.S.: PHLAME: a physical layer aware MAC protocol for
electromagnetic nanonetworks in the Terahertz Band. Nano Commun. Netw. 3(1), 74–81
(2012)
16. Jornet, J.M., Akyildiz, I.F.: Information capacity of pulse-based Wireless Nanosensor
Networks. In: 2011 8th Annual IEEE Communications Society Conference on Sensor, Mesh
and Ad Hoc Communications and Networks, pp. 80–88 (2011)
17. Jornet, J.M., Akyildiz, I.F.: Low-weight channel coding for interference mitigation in
electromagnetic nanonetworks in the Terahertz Band. In: 2011 IEEE International
Conference on Communications (ICC), pp. 1–6 (2011)
18. Wang, P., Jornet, J., Malik, A., Akkari, N., Akyildiz, I.F.: Energy and spectrum-aware MAC
protocol for perpetual wireless nanosensor networks in the Terahertz Band. Ad Hoc Netw.
11(8), 2541–2555 (2013)
19. Piro, G., Abadal, S., Mestres, A., Alarcón, E., Solé-Pareta, J., Grieco, L.A., Boggia, G.:
Initial MAC exploration for graphene-enabled wireless networks-on-chip. In: Proceedings of
ACM NANOCOM, Atlanta, GA (2014)
20. Pierobon, M., Jornet, J., Akkari, N., Almasri, S., Akyildiz, I.F.: A routing framework for
energy harvesting wireless nanosensor networks in the Terahertz Band. Wirel. Netw. 20,
1169–1183 (2013)
21. Liaskos, C., Tsioliarido, A.: A promise of realizable, ultra-scalable communications at nano-
scale: a multi-modal nano-machine architecture. IEEE Trans. Comput. 64, 1282–1295
(2015)
22. Tsioliaridou, A., Liaskos, C., Ioannidis, S., Pitsillides, A.: Lightweight, self-tuning data
dissemination for dense nanonetworks. Nano Commun. Netw. 8, 2–15 (2015)
23. Tsioliaridou, A., Liaskos, C., Ioannidis, S., Pitsillides, A.: CORONA: a coordinate and
routing system for nanonetworks. In: NANOCOM (2015)
The 1st International Workshop on
Multi-clouds and Mobile Edge
Computing (M2EC-2019)
Situation Detection on the Edge
Nikos Papageorgiou1, Dimitris Apostolou1,2(&), Yiannis Verginadis1,

Andreas Tsagkaropoulos1, and Gregoris Mentzas1
1
National Technical University of Athens, 9 Iroon Polytechniou str.,
157 80 Zografou, Athens, Greece
{npapag,jverg,atsagkaropoulos}@mail.ntua.gr,
dapost@unipi.gr
2
University of Piraeus, 80 Karaoli & Dimitriou str., 185 34 Piraeus, Greece
gmentzas@mail.ntua.gr
Abstract. Situation Awareness in edge computing devices is necessary for

detecting issues that may hinder their computation capacity and reliability. The
Situation Detection Mechanism presented in this paper uses Complex Event
Processing in order to detect situations where the edge infrastructure requires an
adaptation. We designed the Situation Detection Mechanism so as it is modular
and can be easily deployed as a Docker container or a set of Docker containers.
Moreover, we designed it to be independent of Complex Event Processing
libraries and we have shown that it can operate with both the Siddhi and Drools
libraries. We evaluated our work with a real-world scenario indicative of the
usage of our component, and its capabilities.
Keywords: Situation Awareness Complex Event Processing

Edge computing
1 Introduction
Mobile Edge Computing (MEC) enables a computing and storage infrastructure pro-
visioned closely to the end-users at the edge of a cellular network. Combining MEC in
multi-cloud infrastructures can help to combat latency challenges imposed by cloud-
centric architectures. However, edge devices are highly dynamic in nature: they are not
as reliable as server and cloud computing resources; they computing capacity is limited
and varies greatly depending on their workload; their operating environment (tem-
perature, humidity, etc.) may impact their performance; they often include sensors
which send data at a very high rate which can sometimes swamp the available network
bandwidth. Perception of these elements in the environment of edge devices within a
volume of time and space and the comprehension of their meaning is typically referred
to as ‘situation’ [6, 7]. Situation detection can enhance the capacity to manage edge
resources effectively as part of a computing environment.
Situations in edge computing infrastructures are highly related to the current status
and context of edge devices and behavior of deployed applications. Situations are in a
rich structural and temporal relationship, and they are dynamic by definition, continu-
ously evolving and adapting. To cope with the dynamicity of situations, one needs to

https://doi.org/10.1007/978-3-030-15035-8_97
986 N. Papageorgiou et al.
sense and process data in large volumes, in different modalities [18]. To realize systems
for Situation Awareness (SA) “individual pieces of raw information (e.g. sensor data)
should be interpreted into a higher, domain-relevant concept called situation, which is an
abstract state of affairs interesting to specific applications” [13]. The power of using
‘situations’ lies in their ability to provide a simple, human-understandable representation
of, for instance, sensor data [13]. In the context of dynamic computing systems, a
situation is defined as an event occurrence that might require a reaction [1]. In com-
puting, SA is the capability of the entities of the computing environment to be aware of
situation changes and automatically adapt themselves to such changes to satisfy user
requirements, including security and privacy [17]. SA is one of the most fundamental
features to support dynamic adaptation of entities in pervasive computing environments.
The research objective of our work has been to design and develop a Situation
Detection Mechanism (SDM) capable of processing in real-time events generated by
the edge infrastructure and detecting situations where the edge infrastructure requires
an adaptation. Moreover, the SDM should be modular and easily deployable on
heterogeneous edge infrastructures. The deployment flexibility of SDM is quite
important since it allows using it on as many as possible different types of edge devices.
SDM is part of the PrEstoCloud dynamic and distributed software architecture that
manages cloud and fog resources proactively, while reaching the extreme edge of the
network for an efficient real-time BigData processing (http://prestocloud-project.eu/).
The rest of the paper is organized as follows. Section 2 discusses enabling tech-
nologies and works related to situation detection and awareness for enterprise decision
making with an emphasis on supporting proactivity. Section 3 outlines the proposed
situation model and described the approach followed to realise situation detection.
Section 4 presents a scenario in which SDM has been evaluated. Section 5 discusses
the main findings of our work and our future plans.
2 Related Work
Situation detection is mainly accomplished using two approaches: specification- and

learning-based ones. Specification-based approaches represent expert knowledge in the
form of logic rules based on event and sensor data, and apply reasoning engines to infer
proper situations from current sensor input [18]. Existing approaches range from earlier
attempts in first-order logic [16] to complex event processing and more advanced logic
models that aims to support efficient reasoning while keeping expressive power, see,
e.g., [14]. With their powerful representation and reasoning capabilities, ontologies
have been widely applied, see [3]. As more and more sensors are deployed in real-
world environments for a long-term experiment, the uncertainty of sensor data starts
poses the need for probabilistic techniques [5] capable of coping with incompleteness,
accuracy, timeliness, and reliability of sensor data [4, 9, 12].
Learning-based techniques have been widely applied to learning complex associa-
tions between situations and sensor data [18]. Typically, learning-based techniques use
supervised learning methods. Supervised learning methods train models using available
labeled data. Manual labelling of training data can be cumbersome and laborious in
cases of many situations and corresponding data to be used for training [10].
Situation Detection on the Edge 987
Unsupervised learning methods can help ameliorate this challenge but there are limited
works reported in the literature. Learning-based methods can cope well with uncertainty
when trained with noisy real-world data [18].
3 Situation Model and Situation Detection Approach
3.1 Situation Model

We follow an event-based approach for situation modeling and detection. We consider
sensor data or event encompassing raw (or minimally processed) data retrieved from
both physical sensors and ‘virtual’ sensors observing systems, services and applications
such as network traffic. We define a situation as an external semantic interpretation of
events. Interpretation means that situations assign meanings to events; external means
that the interpretation is performed from the perspective of applications, rather than
from events; semantic means that the interpretation assigns meaning on events based on
structures and relationships within the same type of events and between different types
of events [18]. A situation can uncover meaningful correlations between events,
labeling them with a descriptive name. The descriptive name can be called a descriptive
definition of a situation, which is about how a human defines a state of affairs in reality.
The PrEstoCloud Situation Metamodel (Fig. 1) captures the concepts based on
which the SDM will be able to detect situations, which may reveal impending failures
or even opportunities for increasing the performance of the deployed applications over
multi-cloud and edge resources. The PrEstoCloudSituation comprises AtomicSitua-
tions and CompositeSituations. An AtomicSituation represents any basic situation
whose value is directly derived from the value of a ComplexEvent. A ComplexEvent is
composed of SimpleEvents (e.g. raw incoming events) and expresses a Scalabil-
ityRequirement (e.g. if RAM > 80% and CPU > 60% for at least 5 min…) that should
drive the Adaptation of the big data intensive application according to a Scalabil-
ityAction (e.g. … then scale horizontally).
The CompositeSituation represents complicated situations pertained to the logical
composition and temporal composition of AtomicSituations. The logical composition
over other situations refer to the ConjunctionSituation (i.e. combining two or more
AtomicSituations using the logical AND operator), DisjunctionSituation (i.e. com-
bining two or more AtomicSituations using the logical OR operator), and Nega-
tionSituation (i.e. combining two or more AtomicSituations using the logical NOT
operator); the temporal composition can be implemented using the TemporalSituation
that describes certain time-related dependencies or sequence associations between two
or more AtomicSituations. A situation may occur before, or after another situation, or
interleave with another situation.
A CompositeSituation can be decomposed into a set of smaller situations, which is
a typical composition relation between situations. For example, a ‘Cold VM migrating’
situation is composed of a ‘Relocating configuration and storage files’ situation, a
‘Moving VM to new host’ situation and a ‘Powering off VM’ situation. According to
our metamodel aggregating SimpleEvents and ComplexEvents we acquire the related
MonitoringInformation which is necessary for checking the health status and QoS of
Fig. 1. PrEstoCloud situation metamodel
both the deployed big data intensive application and the underlying multi-cloud and
edge resources. Thus, all MonitoringInformation is based on the Processing, Host-
ingEdgeNodes, HostingCloudNodes and current Workload detected through the
appropriate software, hardware and workload related monitoring probes, respectively.
Both Processing and Workload are expressed based on the BigDataVocabulary in order
to abstractly map types of big-data streams to big data processing services types
revealing their importance for the detected PrEstoCloudSituations.
3.2 Situation Awareness Approach

In industry, cloud platforms that support automatic or semi-automatic adaptation use
event driven rules in order to decide the time of adaptation. Amazon AWS, for
example, provides auto-scaling services [2] that trigger adaptation actions based on
user-configurable rules that are evaluated in real-time using internal or external mon-
itoring infrastructure. Kubernetes [11] provides auto-scaling capabilities based on
internal or external metrics. In Google Cloud [8], users can specify a target CPU
utilisation for a group of (service) instances, the platform will try to maintain it by
scaling it up or down. OpenStack [15] also supports auto-scaling policies by deploying
the Heat service. Autoscaling in OpenstackHeat is triggered by Alarms produced by the
telemetry service (Ceilometer).
Since a MEC environment combines multi-cloud and edge resources, we need a
mechanism to detect situations from heterogeneous devices and services with very
different capabilities in terms of computational resources and provide the ability to
control and customize the execution environment. For example edge devices may have
very low computational resources or a very restricted (due to security reasons)
environment for custom applications. Very often those devices have low network
bandwidth, unpredictable disconnections from the network and data transmission spikes
that are caused by external events (such as social events, weather conditions or other). In
this environment we need infrastructure and mechanisms for data-driven event detec-
tion. Therefore, we opted for an approach that relies on Complex Event Processing
technologies, which are capable of processing in real-time a large number of events
generated by a variety of distributed cloud and edge computing resources as well as
other data generating sensors. A complex event is an event derived from a group of
events using either aggregation of derivation functions. Information enclosed in a set of
related events can be represented (i.e., summarized) through such a complex event.
Fig. 2. Situation detection mechanism architecture
Arguably, situation detection in a MEC environment needs to take care of network

bandwidth consumption. Similarly, to commercial systems, it is important to support
parts of the situation detection at the edge. This way we can lower resource consumption
in the cloud, limit the required bandwidth or process events from edge devices with
lower latency and lower rates of event loss (due to network outages at the extreme edge).
So, it is crucial for the situation detection mechanism to have low computation resource
consumption (memory and CPU) and ability to efficiently distribute and process events
in multiple stages. The approach that we propose for realising situation detection has the
following characteristics: (i) A homogenous solution for data intensive and data-driven
situation detection at the edge or near the edge and in the cloud. (ii) Components
(containers) that can be deployed with existing cloud orchestration technologies (such as
Kubernetes (https://kubernetes.io/), Rancher (https://rancher.com/), Ansible (https://
www.ansible.com/). (iii) A distributed hierarchical approach for event-driven and rule-
based situation detection with complex event patterns.
The primary input for Situation Detection Mechanism (Fig. 2) is considered any
health status event or application performance-related event transmitted to the Com-
munication and Message Broker (RabbitMQ). Such events are exploited by SDM in
order to reveal problematic situations with respect to the state of the cloud and edge
resources used for hosting big data-intensive applications or to the performance state of
the application itself. Events from the cloud infrastructure (physical and virtual
machines, containers, applications, services, etc.) and edge devices (mobile phones,
routers, IoT devices) are published as events to the Broker in specific topics. One or
more SDM service instances subscribe to the desired topics and receive streams of
events that contain up-to-date information about the current state of those entities (e.g.
used RAM, CPU consumption, disk I/O, requests per second, etc.). SDM instances
process these events based on the supplied CEP rules which are defined in order to
detect interesting situations. Several SDM instances can be used in parallel or in series
in order to process the incoming event streams. High level situations can be detected by
processing low-level situations from many SDM instances.
4 Evaluation
We have evaluated SDM using two different CEP engines, Drools and Siddhi. The
software and hardware configuration is as follows: Hardware (KVM Virtual Machine
with 4 cores and 8 GB RAM running on a server with Intel Xeon E7 @ 2.4 GHz CPU),
Software (SDM services run under Ubuntu 17.10, Docker version 17.12.0-ce, build
c97c6d6, Docker-compose version 1.19.0, build 9e633ef, OpenJDK Runtime Envi-
ronment, Siddhi version v4.0.0 with RabbitMQ extension v1.0.14, Drools version
6.5.0. Final, RabbitMQ 3.7.5 (Docker image rabbitmq3.7.5-management). We used the
RabbitMQ load-testing tool to generate and publish events (https://github.com/
rabbitmq/rabbitmq-perf-test). We used only RabbitMQ and two SDM instances, one
implemented with the Drools CEP library and one implemented with the Siddhi library.
With PerfTest we can select the number of event producers, the length of the period
that we want to send events, the frequency with which the event producers should
generate events and the payload of the events (from a list of files). The AMQP
exchange name and the topic are also configurable. With a Java Management Exten-
sions (JMX) tool such as JConsole or VisualVM (https://visualvm.github.io/), we
monitored the Drools and Siddhi version of SDM.
We ran (with docker-compose) one Siddhi CEP engine and one Drools CEP engine
in parallel and configure them to subscribe to the same AMQP exchange and topic. In
this way both CEP engines received the same events from PerfTest. First we run
PerfTest for 60 s with increasing number of event producers that send one event per
second. The payload of the events is a JSON file that contains different values of two
attributes named “memory” and “cpu”, (without any timestamp for simplification
reasons). Both Drools and Siddhi were configured to produce every 10 s two events
containing: (i) the average CPU and MEMORY (during the last 10 s); (ii) the number
of MEMORY and CPU events that it received (during the last 10 s).
Fig. 3. SDM load testing with PerfTest. CPU utilization of Drools-based implementation vs
Siddhi-based implementation (500, 500, 1000 events/s)
We compared the CPU consumption of Drools and Siddhi while sending 500
events per second (twice) and 1000 events per second, for two consecutive periods of
60 s. We can see clearly in Fig. 3 that the Siddhi-based implementation of SDM has
much lower total CPU utilization than the Drools-based implementation which
increases in a bigger proportion as the rate of incoming events increases.
Fig. 4. SDM load testing with PerfTest. CPU utilization of Drools-based implementation vs
Siddhi-based implementation (1250 to 3000 events/s)
In Fig. 4, we continue the same experiment with increasing number of events per
second (generated by PerfTest): 1250, 1500, 1750, 2000, 2250, 3000. It is again clear
that Siddhi has much lower CPU utilisation. It is also notable that after 1500 events per
second the Drools-based implementation of SDM queues the incoming messages and
continues processing an increasing number of seconds after PerfTest has finished

sending events. Siddhi processes all the events in almost real-time in the above tests.
In Fig. 5, we can see the memory consumption of Drools and Siddhi when sending
500 events per second (twice) and 1000 events per second. In these event rates, both
CEP engines have similar memory consumption. After 1500 events per second Drools
needs more memory than Siddhi (the peak of difference is about 500 MB). If we test
Drools and Siddhi for bigger time periods, over 1500 events per second we can see that
Drools takes much more time to process the incoming events (figures not included for
brevity).
Fig. 5. SDM load testing with PerfTest. Used memory of Drools-based implementation vs
Siddhi-based implementation (500, 500, 1000 events/s)
Further, we compared the expressivity of the Drools and Siddhi rule languages in
expressing situations. Siddhi rule language is SQL-like while Drools language follows
the paradigm of Event-Condition-Action (ECA) rules. Both rule languages have dif-
ferent but straightforward expressions for dealing with simple functions like the
average, the minimum or the maximum of a metric over a period of time. When we
want to aggregate over different objects and group the results over a group key,
according to our knowledge, Drools rule language is not as expressive as Siddhi rule
language. In the following examples we show how we can detect a situation where
more than 20 requests for the same URL are detected in a period of time of 10 s. In
Drools we have to produce an event (RequestAlert) which expires in 10 s and insert it
in the fact base in order to prevent Drools to continuously produce the situation after
the first time it was detected.
/* Drools rule to alert when more than N=20 requests for the same url
are detected in a period of 10 seconds.*/
declare RequestAlert
@role( event )
@timestamp( ts )
@expires ( 10s )
ts: Date @key
url: String @key
end
rule "request_url alert"
timer ( int: 10s 10s )
when
$e1 : SquidEvent( $url : request_url )
not RequestAlert( url == $url )
$c : Number(intValue > 20 ) from accumulate (
$e2 : SquidEvent( this != $e1, request_url == $url ) over
window:time(10s), count( $e2 ) )
then
insert ( new RequestAlert( new Date(), $url ) );
log( "[REQURL] " + $c + " requests to the same url in 10s" );
end
With Siddhi we can group by the attribute “rurl” (that corresponds to the request
URL) and check every 10 s (with #window.timeBatch) for each one if the total is over
20 (having cnt > 20).
/* Siddhi rule to alert when more than N=20 requests for the same url
are detected in a period of 10 seconds.*/
from squidStream#window.timeBatch(10 sec)
select count(rurl) as cnt , rurl
group by rurl
having cnt > 20
insert into rurlStream;
from rurlStream
select str:concat("[REQURL] " , cnt, " requests to the same url
in 10s" ) as msg
insert into msgStream;
5 Conclusions
SDM has been designed and developed focusing on providing detection capabilities for
situations pertaining to edge resources computing capabilities. In implementing SDM,
we followed a specification-based approach. We designed the SDM component so as it
is modular and can be easily deployed as a Docker container or a set of Docker
containers. Moreover, we designed SDM to be independent of CEP libraries and we
have shown that it can operate with both the Siddhi and Drools CEP libraries. The
deployment flexibility of SDM is quite important since it allows using the Complex
Event Processing engine of choice based on the processing capabilities required in each
case and the prior expertise regarding a certain Complex Event Processing engine.
We evaluated SDM with a real-world scenario indicative of the usage of our

component, and its capabilities. Testing and evaluation of SDM revealed that it is
capable to detect situations defined as complex event patterns. Specifically, we tested
SDM in conjunction with both Drools and Siddhi in two scenarios: first, we stress-
tested it using the PerfTest load-testing tool of RabbitMQ and, second, to detect sit-
uations in computer network traffic in a real production computing environment. Tests
indicated that SDM can be used to detect situations expressed as complex event pat-
terns. Moreover, out tests have shown that Siddhi can scale better than Drools.
In the future, we will further test SDM with more complex patterns and scenarions.
Moreover, we will augment the specification-based approach we followed for SDM
with learning-based methods and techniques to cope with more and more complex
situations, which cannot be manually specified as well as with imperfect sensors.
Moreover, our future work will focus on enhancing SDM with capabilities to recom-
mend adaptations to the edge processing topology in order to optimize their usage so
that its performance requirements can be satisfied.
Acknowledgments. This work is partly funded by the H2020 PrestoCloud project–Proactive

Cloud Resources Management at the Edge for Efficient Real-Time Big Data Processing
(732339).
References
1. Adi, A., Etzion, O.: Amit - the situation manager. VLDB J. 13(2), 177–203 (2004)
2. Amazon (2018). https://aws.amazon.com/autoscaling
3. Chen, H., Finin, T., Joshi, A.: An ontology for context-aware pervasive computing
environments. Knowl. Eng. Rev. 18(3), 197–207 (2004)
4. Cohen, N.H., Lei, H., Castro, P., Davis II., J.S., Purakayastha, A.: Composing pervasive data
using iQL. In: Proceedings of the Fourth IEEE Workshop on Mobile Computing Systems
and Applications, WMCSA 2002, pp. 94–104 (2002)
5. Haghighi, P.D., Krishnaswamy, S., Zaslavsky, A., Gaber, M.M.: Reasoning about context in
uncertain pervasive computing environments. In: Proceedings of the 3rd European
Conference on Smart Sensing and Context, EuroSSC 2008, pp. 112–125. Springer,
Heidelberg (2008)
6. Endsley, M.: Designing for Situation Awareness: An Approach to User-Centered Design,
2nd edn. CRC Press, Boca Raton (2016)
7. Franke, U., Brynielsson, J.: Cyber situational awareness - a systematic review of the
literature. Comput. Secur. 46, 18–31 (2014)
8. Google Cloud (2018). https://cloud.google.com/compute/docs/autoscaler/
9. Gray, P.D., Salber, D.: Modelling and using sensed context information in the design of
interactive applications. In: Proceedings of the 8th IFIP International Conference on
Engineering for Human-Computer Interaction, EHCI 2001, pp. 317–336. Springer, London
(2001)
10. Gu, T., Chen, S., Tao, X., Lu, J.: A non supervised approach to activity recognition and
segmentation based on object-use fingerprints. Data Knowl. Eng. 69(6), 533–544 (2010)
11. Kubernetes (2018). https://kubernetes.io/docs/tasks/run-application/horizontal-pod-auto-
scale-walkthrough/
12. Lei, H., Sow, D.M., John, S., Davis, I., Banavar, G., Ebling, M.R.: The design and applications
of a context service. SIGMOBILE Mob. Comput. Commun. Rev. 6(4), 45–55 (2002)
13. Loia, V., D’Aniello, G., Gaeta, A., Orciuoli, F.: Enforcing situation awareness with granular
computing: a systematic overview and new perspectives. Granular Comput. 1(2), 127–143
(2016)
14. Loke, S.W.: Incremental awareness and compositionality: a design philosophy for context-
aware pervasive systems. Pervasive Mob. Comput. 6(2), 239–253 (2010)
15. OpenStack (2018). https://docs.openstack.org/senlin/latest/scenarios/autoscaling_heat.html
16. Ranganathan, A., Al-Muhtadi, J., Campbell, R.H.: Reasoning about uncertain contexts in
pervasive computing environments. IEEE Pervasive Comput. 03(2), 62–70 (2004)
17. Yau, S.S., Liu, J.: Hierarchical situation modeling and reasoning for pervasive computing.
In: The Fourth IEEE Workshop on Software Technologies for Future Embedded and
Ubiquitous Systems, SEUS 2006/WCCIA 2006, pp. 6-pp, April 2006
18. Ye, J., Dobson, S., McKeever, S.: Situation identification techniques in pervasive
computing: a review. Pervasive Mob. Comput. 8(1), 36–66 (2012)
A Context-Aware Service for Authorizing
Multi-cloud Deployments
Yiannis Verginadis(&), Ioannis Patiniotakis(&),

and Gregoris Mentzas(&)
Institute of Communications and Computer Systems,

National Technical University of Athens,
9 Iroon Polytechniou Str., Athens, Greece
{jverg,ipatini,gmentzas}@mail.ntua.gr
Abstract. The most recent advancements in cloud computing, highlight the

need for supporting deployments on virtualised resources bounded to the data-
intensive application requirements and not limited by the strict boundaries of
each available cloud provider. Towards this direction, one important challenge is
the appropriate protection of the software infrastructure used for automatically
performing application components deployment over multi-clouds. Specifically,
we discuss the requirements and introduce a novel attribute-based access control
mechanism, able to cope with potential cybersecurity threats that may com-
promise the deployment of multi-cloud applications. We attempt to tackle the
authorization issues from two different perspectives; namely, coping with the
“access control” to various platform components and the “pre-authorization” of
application deployment and data placement actions using multiple cloud
providers.
1 Introduction
In recent times, there is an abundance of cloud services offered from a vastly increasing
number of cloud providers, each one with its own advantages or shortcomings. Finding
opportunities where the benefits of each provider are exploited at the same time, while
coping with the ever-increasing requirements of the modern data-intensive cloud
applications, has become imperative [1]. A generic challenge is to overcome scala-
bility, resiliency, and security issues faced by big data and data-intensive applications
on distributed platforms by using transparent and optimized multi-cloud resource
provisioning [1]. One of the critical issues towards tackling this generic challenge is to
design and develop the appropriate methods and tools for adequately protecting the
platform components that make possible the optimized application placement on multi-
clouds. We use as an example the platform presented in [2] where decision and
optimization components (called Upperware) are used for ingesting user-defined
application placement requirements, constraints and optimization goals in order to
provide optimized decisions on initial placement or reconfigurations of data-intensive
applications on multi-clouds. A second part of such a platform involves a set of
components that interface with the required providers for commissioning virtual
resources and deploying application components according to the initial placement or

https://doi.org/10.1007/978-3-030-15035-8_98
A Context-Aware Service for Authorizing Multi-cloud Deployments 997
the reconfiguration decisions of the Upperware components. This platform [2] follows
a model-driven engineering approach, where models as first class citizens, initially
capture placement requirements and optimization goals, providing valuable input for
analysis and decision making with respect to multi-cloud application placement and
last they are transformed to executable service graphs that guide the placement or
reconfiguration actions across clouds.
Nevertheless, the list of the main risks associated with clouds is very long, and even
longer for multi-clouds [3]. At the top of this list, we find the improper identity
management, insufficient credentials and access control, unsecured application pro-
gramming interfaces (APIs) [4]. Although there are several security-related topics that a
platform that undertakes applications placement on multi-clouds should cover (e.g.
cloud providers credentials management, component authentication, encrypted com-
munication etc.), in this work we are focusing on an advanced authorization module,
based on the Attribute-based Access Control (ABAC) paradigm. This mechanism is
designed to provide access control and pre-authorization of any deployment plans that
may be implemented by such a platform. We note that such security related capabilities
are applicable to any application placement scenario but we mainly focus on data-
intensive applications due to the inherent additional requirements for efficiency that
introduce. The introduction of this kind of authorization capabilities, ensures that only
eligible entities (users or components) can access protected platform resources and
apply certain operations on them. Each access attempt to a resource is checked against
a set of access control policies. During authorization checking, various stated and
contextual information must be considered; this information relates to the requestor (i.e.
user or component), the resource being accessed (e.g. data, methods etc.), the attempted
operation as well as other environment data (e.g. date, component operational status,
etc.). Beyond access control, the authorization infrastructure should be consulted on
whether a given application deployment plan, generated by the appropriate decision
and optimization components, complies with a set of deployment policies. Such
policies may encompass constraints and limitations referring to application deployment
(e.g. total cost or number of virtual machines deployed or location of resources). In this
way, the described security enhancement of the platform leads to a protected operation
of all critical software that makes decisions and implements placement or reconfigu-
ration actions regarding multi-cloud deployments.
The rest the paper involves a relevant state-of-the-art analysis which is provided in
Sect. 2, while in Sect. 3, we discuss the requirements of the developed service called
Melodic Authorization Service. In Sect. 4, we provide the design details of this
mechanism, while in Sect. 5 the implementation specifics of this service are described.
Last, we conclude with Sect. 6, where the next steps of this work are discussed.
2 Related Work on Access Control
Several access control models have been proposed in the literature and used in software
products. These models provide a framework and a method of how resources,
requestors, operations, and rules may be combined to produce and enforce an access
control decision. Some of the most well-known models are the Discretionary Access
998 Y. Verginadis et al.
Control (DAC) [5], the Mandatory Access Control (MAC) [6], the Identity-Based
Access Control (IBAC) [6], the Role-Based Access Control (RBAC) [7] and the Attri-
bute-Based Access Control (ABAC) [7]. Out of the most prominent paradigms is the
ABAC, especially if we consider the dynamic characteristics of the cloud computing
domain. ABAC uses policies that comprise rules, which in turn comprise logical
conditions on several different attributes. Typically, each rule contains at least a con-
dition (i.e. boolean expression) and a decision (permit or deny) to reach when the
condition is true. Policies combine the outcomes of rules and yield the final decision
using certain outcome combination methods. Attributes can be properties of the
requesting entity, of the resource being accessed, of the operation requested, or any
other contextual information found in the so-called external environment enclosing the
interacting entities. Thus, these attributes are not limited in number or in type as in
previous paradigms (e.g. RBAC which only considers the role of the access requesting
entity).
Table 1. Java-based, open source, XACML tools

Product XACML License Latest release Findings
version
WSO2 Balanaa 3.0, 2.0, 1.x Apache 2.0 Mar 2018 Based on Sun’s XACML
The most used XACML
implementation so far
Authzforce CEb 3.0 Apache 2.0 Apr 2018 Lack of clear documentation
for extensions development
JBoss Picketboxc 2.0 LGPL 2.1 Feb 2011 Merged with Keycloak project
since 2015
Xacml4jd 3.0, 2.0 GPL 3.0 Jul 2014 No recent activity at Github
since 2014
XACML Lighte 2.0 n/a Apr 2013 PDP & PAP only
Heras AFf 2.0 Apache 2.0 Aug 2016 Basic XACML 2.0
implementation
OpenAZg 3.0 Apache 2.0 Mar 2016 Retired since Aug 2016
Sun’s XACMLh 2.0, 1.x Open source Dec 2010 No active support anymore
a
https://github.com/wso2/balana
b
https://authzforce.ow2.org/
c
http://picketbox.jboss.org/
d
https://github.com/xacml4j-opensource/xacml4j.github.io
e
http://xacmllight.sourceforge.net/
f
https://bitbucket.org/herasaf/herasaf-xacml-core
g
http://incubator.apache.org/projects/openaz.html
h
http://sunxacml.sourceforge.net/
There are a few reference implementations of the ABAC model, but among the
most important ones are the eXtensible Access Control Markup Language (XACML)
and the Next Generation Access Control (NGAC) [8]. It is important to note that
XACML seems to be the most widely used ABAC implementation model, since it
enjoys worldwide industrial adoption in sectors like banking, healthcare, and insurance
[9]. XACML is an XML-based, open-standard language promoted by OASIS, for
expressing authorization policies and querying access to resources. Evaluating an
access request to a resource, with regards to an XACML policy, may result in one of
these four values: Permit, Deny, Indeterminate (an error occurred or needed values
were missing) or Not Applicable (no related policy found). The XACML specification
defines five main components that handle access decisions; the Policy Enforcement
Point (PEP), Policy Administration Point (PAP), Policy Decision Point (PDP), Policy
Information Point (PIP), and a Context Handler (CH) [8]. Since XACML introduction,
several compliant frameworks have been developed and offered, both as open source
software (Table 1) as well as proprietary products.
We note that we have thoroughly discussed the numerous limitations of previous
works that do not follow the ABAC paradigm or its reference implementations for
coping with the challenges of data authorization in dynamic multi-cloud environments
[4]. In addition, this work proceeds further than previous efforts that try to overcome
the basic syntactic descriptions by capturing the knowledge that lurks behind policies
and rules in the sense that the relations between the attributes used for access control
are considered [4]. To the best of our knowledge such a context-aware access control
mechanism has not been implemented for securing the model-driven engineering
interactions, within platform components that enable multi-cloud deployments and
reconfigurations. Such a mechanism aspires to alleviate the repercussions of any
potential components’ compromise, coming from external adversaries, by authorizing
in a context-aware manner any updates on the models used.
3 Melodic Authorization Service Requirements
Authorization refers to a security mechanism that determines and enforces access

privileges of a requesting entity, related to resources and application features. In our
approach, this service materialises two objectives. First, the supply of a security-by-
design access control framework for protecting all the critical platform components that
undertake the task of deploying application components over several multi-cloud
resources. A platform that undertakes the decision-making and control of the way that
applications can be deployment in multi-clouds, comprises a set of network-connected
microservices, distributed over an intranet or a (virtual) private network. Despite the
significant advantages of this approach, certain attack vectors exploiting the networked
and distributed nature of the platform are possible (e.g. database-related and denial-of-
service attacks). Furthermore, we consider the semi-honest adversarial model, where a
malicious cloud provider can intercept all messages and may attempt to exploit them in
order to learn information that otherwise should remain private. In order to ensure a
sufficient level of security, it is necessary to protect platform components from
unauthorized internal or external access attempts. For this reason, additional parameters
must be taken into account; for instance, the components’ previous behaviour, the
origin and time of an access request, or the current state and environment of the
platform. Such information is usually termed as context [10]. Contextual information
can be of various types and originate from diverse sources. The Metadata Schema
introduced in [1], provides a classification of these information types, in its Context-
aware Security model (e.g. permission types, physical or network location of an
incoming request, time-related patterns based on normal access requests, etc.). This
classification can act as a common vocabulary for collecting and leveraging informa-
tion for authorization purposes.
The second objective relates to the enforcement of policies and limitations
regarding the deployment of cloud applications and their data across various cloud
providers. By design, a decision-making component in Upperware is responsible for
producing correct deployment plans conforming to any given constraints and limita-
tions. However, a compromised component, due to a cyber-attack, could possibly yield
invalid deployment plans. Therefore, a precautionary validation step, before the actual
deployment, would mitigate the risk of deploying an application in a non-conformant
manner. We refer to this step as pre-authorization. The enforced limitations can be
regulatory (e.g. data not allowed to leave the EU), corporate, as well as budget-,
resource- or security-related. A pre-authorization policy could for example, pose a limit
on the number of virtual machines deployed on a cloud provider, or require the storage
of data of a certain type to be stored in nodes located only in the EU. These two
objectives are quite different in their business purpose and involve different autho-
rization rules. However, the same authorization capabilities and toolset can be used in
order to achieve both of them.
4 Melodic Authorization Service Approach
In this section the details of the developed authorization service approach are high-
lighted. Specifically, we discuss the conceptual architecture of the system and provide
details on how the contextual attributes are managed by this system.
It is important to note that we have adopted the ABAC model for this service,
according also to the analysis provided in Sect. 2. Specifically, based on the infor-
mation provided in Table 1, we opted to use WSO2 Balana engine for XACML 3.0
implementation. However, replacing it with another alternative is expected to be a
relatively straightforward task, since the XACML policy engine resides inside the PDP
component of the XACML architecture. Furthermore, the pluggable design of the
server will allow the easy replacement of plugins pertaining to the specific policy
engine with new ones.
4.1 Authorization Service Architecture

Figure 1 depicts the architecture of the developed Authorization Service that enhances
the XACML reference architecture [9]. The server part of the service that receives and
evaluates access requests is enclosed in a dashed box in the same figure. We consider
one PEP embedded within each of the multi-cloud management platform components
that must be protected. This is where incoming access requests to component resources
enter the platform. A PEP intercepts any incoming request, interrupts the normal
request flow, extracts request information and then contacts the Authorization Server
relaying the extracted information. If the server returns a positive decision, the access
request processing flow resumes. Otherwise, the access is prevented. PEP is provided
as an authorization service client library, which is embedded in the platform compo-
nents that should be protected. The communication to PDP is achieved using the
REST API exposed by the Authorization Server, over an encrypted transport layer
security (TLS) connection.
These PEPs communicate through an appropriate load balancing to a certain
Authorization Server. Depending on the use case, multiple Authorization Servers can
be used for scaling the access control capabilities that are provided by our system.
Specifically, an Authorization Server comprises a PDP, a Context Handler and several
PIPs. Each PDP constitutes a web service that provides a RESTful API for receiving
access request information from PEP’s, evaluating them against policies and eventually
authorizing or declining the certain access request. For this purpose, PDP contains a
policy evaluation engine, namely the WSO2 Balana. Upon configuration, PDP will first
invoke the Context Handler to collect additional (contextual) information from the
request or the environment, and then evaluates the incoming request against a number
of pre-defined access control policies. Several PDP nodes may coexist in a cluster to
achieve high availability, fault tolerance and fast response times. Typically, all PDP
nodes share the same configuration and the same policy repository. A second com-
ponent embedded in the Authorization Server is the Context Handler. Upon activation,
it invokes the configured plugins to collect additional information (as attributes) about
the context of the request. This contextual information is subsequently stored in a PIP
(in order to become available during the policy evaluation and access control decision.
Furthermore, the Context Handler receives platform or environment-related context
from Context Collectors. The PIP is responsible for providing values for all the
involved attributes that participate in the access control rules and policies, based on
which an incoming request is permitted or denied. Specifically, the PDP’s policy
evaluation engine, while processing a request, might require attribute values not con-
tained in the request itself. In such case, it invokes PIP plugins to retrieve the required
attributes that constitute key-value pairs, where keys are typically in the form of
Uniform Resource Names (URN).
Fig. 1. Authorization service conceptual architecture

The PAP is implemented in order to manage and store the XACML policies. These
policies are stored in a shared place accessible by all PDPs. The architecture is com-
plemented by a number of Context Collectors (CC), i.e. independent and domain-
specific application components that aim at continuously collecting information about
the multi-cloud management platform and its environment. Thus, the CCs are essen-
tially mechanisms for collecting contextual information from the environment and
propagate it to the Context Handlers. It is expected that different context information,
and thus context collectors, will be needed in different multi-cloud deployments based
on the application domain. Last, appropriate load balancing capabilities have been
fused into our system. Specifically, the PDP Load-Balancing stands between PEP
clients and the PDP nodes to prevent flooding of one or more of the Authorization
Server instances and allowing the proper scaling of the system. This is implemented at
the PEP-side as a dynamically configured list of PDP endpoints that are contacted
either in successive order (round-robin) or selected randomly. Similarly, Context
Handler load balancing capabilities can be used for enabling CCs to efficiently com-
municate with the appropriate CHs. We note that it is also feasible to add a third-party
HTTP proxy or an open source load-balancer (e.g. Traefik1) by configuring the PEP
clients. In either way, incoming access control requests are dispatched in a balanced
way to PDP cluster nodes and the raw context is collected and handled proportionally
by the CHs. We note that any of the incoming requests to be authorized come from
authenticated entities which implies the implementation and integration of appropriate
authentication mechanisms, a work that is out of the scope of this paper.
Fig. 2. Attribute values flow in the authorization service
4.2 Context Attributes in Authorization Service

The attributes handled by the Authorization Service can be of three types; (a) access
request-related attributes (Fig. 2 – Step 1, e.g., requestor id, resource id), (b) request
context attributes in the sense that these are not stated in examined access request but
1
https://traefik.io/.
are acquired from other sources (Fig. 2 – Step 2b, e.g., requestor location and device),
and (c) environment/platform-related context attributes, not pertaining to a specific
access request (Fig. 2 – Step 2a, e.g., operational status of a platform component). The
difference in the context in the two latter cases is that the request context becomes
invalid when the request has been processed, whereas environment/platform context
evolves independently of the access requests. Figure 2 provides a high-level view of
the attribute flow in the Authorization Service. Specifically, once the context is col-
lected (Fig. 2 – Steps 1, 2a, 2b), it becomes available for the PIP (Fig. 2 – Step 3),
which upon request provides them in the form of key-value pairs to the PDP for issuing
an access control decision (Fig. 2 – Step 4).
5 Implementing Authorization Service Approach
In this section, we highlight the key components of a platform dedicated for managing
applications deployments on multi-clouds that can be protected by exploiting the
implemented Authorization Service. This kind of protection is enabled by deploying
appropriate PEPs on each of the following platform components. In this way, we
consider a minimal overhead in the implementation of the developed Authorization
Service, since the approach implies that a PEP library is bundled with the application
code (i.e. same Java Archive (JAR) or placed in the Classpath), while a Tomcat
interceptor or “@AuthorizationRequired” annotations on guarded methods are added.
Business Process Management engine (BPM). It coordinates all the decision and
optimization components (i.e. Upperware components in [2]) and executes the work-
flow to generate and execute an application deployment plan, according to a set of user-
defined preferences, constraints and optimization goals. When necessary, it also repeats
the whole process or parts of it to introduce deployment plan updates, as a response to
changes in application demands or environment. A PEP client has to be embedded in
BPM engine, in order to protect it from a potentially compromised or malfunctioning
component or from outside-world interactions. In this way each PEP examines the
origin and timeliness of the requests in order to authorize them.
Deployment and Adaptation mechanism. It is responsible for taking an application
deployment plan and executing it by providing specific instructions to components able
to interface with several cloud providers that should be used, for commissioning VMs
and installing application components accordingly. In order to verify that a given
deployment plan conforms to the application deployment policies, a pre-authorization
step is taken. The plan parameters are checked against the relevant policies, and if
rendered as conformant, the deployment starts. For this reason, the Deployment and
Adaptation mechanism uses a PEP client to contact PDP to evaluate the plan against
the posed policies. The plan pre-authorization policies are different from access
authorization policies used for checking the access to previous components.
Modelling Editors. They are used to firstly create and maintain the Metadata
Schema [1], valuable for providing a formal hierarchical view of the contextual attri-
butes to be used for modelling the application placement problem. Secondly, it is used
for acquiring, from the DevOps, the appropriate application placement requirements,
constraints and optimization goals that should guide the platform when undertaking the
multi-cloud placement of the given application components. These editors comprise

two layers; the User Interface layer, which executes in user browser, and the Backend,
modelling management layer. The latter also communicates and interacts with models
repository. For this reason, the second layer includes a PEP client to protect itself from
unauthorized access to its functionality and data.
5.1 Use of Request Interceptor for Spring-Boot Based Components

Most the platform components have been implemented as Spring-Boot web applica-
tions2. This means they embed a minimal application server (e.g. Tomcat3) in order to
accept incoming (HTTP) requests from other platform components, providing suitable
REST APIs. The code implementing the REST API and receiving the requests needs to
be protected with this Authorization Service.
One method for introducing the needed authorization capabilities is by configuring
the embedded Tomcat server (of the Spring-Boot framework) to intercept the incoming
requests and pre-process them before they actually reach the code that serves them. This
is a standard step in the Tomcat HTTP request processing cycle and is implemented by
adding special filters called interceptors. Interceptors can be added in Tomcat pro-
grammatically, during server initialisation. The interception process of an HTTP request
is handled by a Login Interceptor. This interceptor is invoked three times: (a) Pre-
Handle: before calling the code that is meant to service the request (i.e. MainController),
(b) Post-Handle: after the MainController returns and before rendering the response, and
(c) After-Completion: when response has been sent back to requestor. Regarding
Spring-Boot web applications, interceptors can be added using an application config-
uration class that implements the WebMvcConfigurer interface. Therefore, all needed
interceptors can be added in the Tomcat interceptor registry, before the server starts.
This approach does not require any modification of application source code.
Instead, a new configuration class can be written to configure an authorization inter-
ceptor. This class must be packaged with existing code, and Spring-Boot will take care
of using it at runtime. The downside of this method is that it applies only to Spring-
Boot web applications using a Tomcat or Jetty server4.
5.2 Use of Aspect and Aspect-Oriented Programming

A second approach for programmatically exploiting the developed Authorization
Service consists of using the Aspect-Oriented Programming (AOP) paradigm. AOP is a
programming approach for software modularization and separation of cross-cutting
concerns [11]. This is achieved by adding extra functionality (called Advice) to
existing code without significantly modifying the source code (i.e. annotations). This
addition typically occurs during the software building phase in a task called weaving,
which is undertaken by specialised tools called weavers. The code to be modified is
2
http://spring.io/projects/spring-boot.
3
http://tomcat.apache.org/.
4
https://www.eclipse.org/jetty/.
identified via pointcuts, which are specifications of those code artefacts (typically
classes and methods) that need to be enhanced with advices. An advice, along with the
pointcuts that specify the code it must be applied onto, is called an Aspect. We note that
the Before aspects are used to enforce PEP functionality, which means that a call to a
guarded method is intercepted, PEP checks that this call is permitted and then the actual
method execution occurs.
AOP allows the non-core functionality of a software component (e.g. logging of
code executions, measuring duration, and authentication/authorization) to be moved
away from the code implementing the core business of the component. The non-core
functionality is added and interleaved with the core functionality during the software
build phase (via weaving). The Spring framework provides an AOP implementation.
Spring AOP is proxy-based, meaning each code artefact that can be enhanced with
advices will be wrapped by a suitable proxy object that is actually invoked by the
calling code. The proxy can subsequently pass control to the actually requested code.
Proxies are automatically introduced at code-level (during weaving), while source code
remains intact. Thus, this process is transparent to the programmer.
Regarding the use of Aspects in the Authorization Service, an authorization aspect
has been introduced. The corresponding advice (i.e. the wrapping proxy code) will
intercept the code invocation in order to perform a series of authorization related tasks;
namely, (a) create/reuse a PEP client object, (b) collect invocation information (i.e.
method signature and arguments), (c) connect to a PDP server and pass the collected
information, (d) receive the PDP server response (permit, deny, error), and (e) in case
of permit (subsequently), call the actual (wrapped) code or raise an authorization error,
otherwise. When the wrapped code is a method of a Web or REST controller class, and
that method is mapped to a Web or REST endpoint, then the corresponding (HTTP)
request object is introspected to extract all related information.
6 Conclusions
In terms of this work, a novel authorization service was proposed in the context of
protecting components that undertake the data-intensive application deployments on
multi-clouds. Specifically, this work is seen from two different perspectives; namely,
the “access control” to various platform components and the “pre-authorization” of
application deployment and data placement in cloud providers. In the first case,
authorization capabilities are considered to be responsible for protecting the platform
itself from illegal access attempts and interference with its normal operation. In the
latter case, authorization capabilities refer to the pre-authorization of application
deployment and data placement plans, produced by the decision-making and opti-
mization components, considering a given set of policies, constraints and goals.
The next steps of this work, involve the appropriate integration with other critical
security services that will address the security challenges in the multi-clouds deploy-
ment domain in a holistic manner. Such additional security mechanisms involve the use
of authentication services, encrypted communication (e.g. SSL/TLS), mechanisms for
counterfeiting the man-in-the-middle attacks, using component verification and digital
certificates among others.
Acknowledgments. The research leading to these results has received funding from the
European Union’s Horizon 2020 research and innovation programme under grant agreement
No. 731664. The authors would like to thank the partners of the MELODIC project (http://www.
melodic.cloud/) for their valuable advices and comments.
References
1. Verginadis, Y., Patiniotakis, I., Mentzas, G.: Metadata schema for data-aware multi-cloud
computing. In: INISTA, pp. 1–9. IEEE (SMC) (2018)
2. Horn, G., Skrzypek, P.: MELODIC: utility based cross cloud deployment optimisation. In:
Proceedings of the 32nd International Conference on Advanced Information Networking and
Applications Workshops (WAINA) (2018). https://doi.org/10.1109/waina.2018.00112
3. CSA: The Treacherous 12 - Cloud Computing Top Threats in 2016. https://
cloudsecurityalliance.org/group/top-threats
4. Verginadis, Y., Michalas, A., Gouvas, P., Schiefer, G., Hübsch, G., Paraskakis, I.:
PaaSword: a holistic data privacy and security by design framework for cloud services.
J. Grid Comput. 1–16 (2017). https://doi.org/10.1007/s10723-017-9394-2
5. Decker, M.: Modelling of location-aware access control rules. In: Maria Cruz-Cunha, F.M.
(ed.), Handbook of Research on Mobility and Computing: Evolving Technologies and
Ubiquitous Impacts, Information Science Reference, Hershey, PA, USA, 2011, incollection
57, pp. 912–929 (2011)
6. Hu, V.C., Ferraiolo, D., Kuhn, R., Friedman, A.R., Lang, A.J., Cogdell, M.M., Schnitzer, A.,
Sandlin, K., Miller, R., Scarfone, K.: Guide to attribute based access control (ABAC)
definition and considerations (draft). NIST Special Publication 800, 162 (2013)
7. Ferraiolo, D., Cugini, J., Kuhn, D.R.: Role-based access control (RBAC): features and
motivations. In: Proceedings of 11th Annual Computer Security Application Conference,
pp. 241–248 (1995)
8. Ferraiolo, D., Chandramouli, R., Kuhn, R., Hu, V.: Extensible access control markup
language (XACML) and next generation access control (NGAC). In: Proceedings of the
ACM International Workshop on Attribute Based Access Control, pp. 13–24. ACM (2016)
9. Garcia-Alfaro, J., Navarro-Arribas, G.: Prevention of cross-site scripting attacks on current
web applications. In: OTM Confederated International Conferences on the Move to
Meaningful Internet Systems, pp. 1770–1784. Springer (2007)
10. Abowd, G.D., Dey, A.K., Brown, P.J., Davies, N., Smith, M., Steggles, P.: Towards a better
understanding of context and context-awareness. In: International Symposium on Handheld
and Ubiquitous Computing, pp. 304–307. Springer (1999)
11. Kiczales, G., Lamping, J., Mendhekar, A., Maeda, C., Lopes, C., Loingtier, J., M., Irwin, J.:
Aspect-oriented programming. In: European conference on object-oriented programming,
pp. 220–242. Springer (1997)
A Real-Time Video Streaming System
over IPv6+MPTCP Technology
Yu Luo1 , Xing Zhou1(B) , Thomas Dreibholz2 , and Hanbao Kuang1

1
College of Information Science and Technology, Hainan University,
Haikou, Hainan, China
18636038948@163.com, xingzhou50@126.com, 1521793003@qq.com
2
SimulaMet, Oslo, Norway
dreibh@simula.no
Abstract. Today, a steadily increasing number of users are not just pas-
sively consuming Internet content, but also share and publish content.
Users publish text, photos and videos. With the availability of 5G high-
speed, low-latency mobile broadband networks, real-time video stream-
ing will also be possible. We believe this will become a very popular
application in the coming years. But the more popular a service is, the
higher the need for resilience. In this paper, we introduce our work-in-
progress live video streaming platform for future mobile edge computing
scenarios, which makes use of MPTCP+IPv6 to support multi-homing
for resilience and multi-path transport for load balancing. As a proof
of concept, we will show that the platform is (1) compatible with IPv6,
(2) utilizes load balancing when possible and (3) provides robustness by
network redundancy.
Keywords: Video streaming platform · IPv6 · MPTCP ·

Load balancing · Resilience
1 Introduction
With the rapid development of the Internet, more and more real-time services
are deployed in the network, such as live video broadcast, video conferencing and
other key real-time businesses. These applications have stricter requirements for
bandwidth, latency and jitter. 5G networks, together with cloud infrastructure
for mobile edge computing (MEC), are a key enabler for such applications. But in
addition, there is also need for network and transport improvements: Today, IPv4
addresses are a scarce resource, compared with new applications like the Internet
of Things (IoT). IPv6 is therefore an important development step of the next
This work has been funded by the NSFC of China (No. 61363008/61662020), CERNET
NGI Technology Innovation Project (No. NGII20160110) and the Research Council of
Norway (Forskingsrådet) (prosjektnummer 208798/F50).
The authors would like to thank Ted Zimmerman for his comments.
https://doi.org/10.1007/978-3-030-15035-8_99
1008 Y. Luo et al.
generation Internet. With the rapid development of the IoT, various network
services and applications based on IPv4 will definitely migrate to IPv6 [26]. But
the single-path transmission patterns of TCP cannot handle multiple addresses
per endpoint or address changes, which are usual for IPv6. Therefore, another
milestone is Multi-Path TCP (MPTCP) [16,24], providing transport redundancy
by multi-homing without changing the existing network infrastructure. It also
supports load balancing over heterogeneous networks connections [15,17,27,30].
But MPTCP is not only useful when an endpoint is connected to multiple Inter-
net connections. [20,22] measured and compared the stability of IPv4 and IPv6
Internet paths. With IPv4/IPv6 dual-stack endpoints, MPTCP may already pro-
vide additional redundancy by just using non-congruent IPv4 and IPv6 paths.
In this paper, we introduce our work in progress on a live video streaming
service based on MPTCP and IPv6, as a proof-of-concept for a resilient service
in MEC. We furthermore present initial evaluation results.
2 Overview of Relevant Technologies

The major limitation of IPv4 is its limited address space: with its 32-bit
addresses, it theoretically provides almost 4.3 billion (232 ) unique Internet
addresses, while only around 3.7 billion addresses are usable host addresses.
IPv6 [3] therefore extends the address space to 128 bits, theoretically allowing
around 340 undecillion (2128 ) addresses. This allows for a very systematic assign-
ment of subnets [26]. IPv6 furthermore provides some improvements for more
efficient handling of packets [2], like an optimized header format.
TCP establishes a connection between two IPv4 or IPv6 addresses. Port
numbers are used to identify the connection, i.e. the 4-tuple (source address/-
port, destination address/port) uniquely identifies the connection. TCP neither
supports multiple address per endpoint nor address changes during connection
lifetime. The MPTCP extension [16] adds this multi-homing feature. Addresses
are negotiated during connection establishment. Furthermore, addresses may
change during connection lifetime. MPTCP supports IPv4 and IPv6; boths pro-
tocols can be used simultaneously in the same connection. Furthermore, MPTCP
can utilize multiple subflows simultaneously (multi-path transport) and balance
load among the paths.
MPTCP establishes so-called subflows, which look – for middlebox devices
like firewalls – like regular TCP flows. MPTCP therefore works in existing net-
works, with existing middlebox devices. That is, unlike for other protocols like
the Stream Control Transmission Protocol with Concurrent Multipath Trans-
fer extension (CMT-SCTP, [5]), it is compatible with old networking devices.
Only the two endpoints of an MPTCP connection need MPTCP support. This
ensures a smooth deployment [1,15,18]. MPTCP provides the same Socket API
interface [19,25] to the Application Layer as standard TCP. That is, applications
interact with MPTCP as they would do with TCP. In most cases, the application
does not even have information about an underlying MPTCP instead of TCP.
A Real-Time Video Streaming System over IPv6+MPTCP Technology 1009
Table 1. Socket API functions and IPv6 communication
Socket API function Functional description

inet pton() Convert address string to address
inet ntop() Convert address to address string
gethostbyname() Translate name/address string to IP
address(es) – deprecated!
gethostbyaddr() Translate IP address to name/address string
– deprecated!
getaddrinfo() Translate name/address string to IP
address(es)
freeaddrinfo() Free memory allocated by getaddrinfo()
getnameinfo() Translate IP address to name/address string
socket() Create socket
bind() Bind socket to address
listen() Put socket into listening mode (TCP, SCTP)
accept() Accept incoming connection (TCP, SCTP)
connect() Establish new connection
send() Send data (TCP, SCTP)
sendto() Send data (UDP)
recv() Receive data (TCP, SCTP)
recvfrom() Receive data (UDP)
close() Close socket
3 Using Software with IPv6 and MPTCP

In the following, we introduce how to adapt systems with IPv6 and MPTCP
support.
3.1 Making Software Compatible with IPv6

Network and Transport Layer protocols are used in applications by the oper-
ating system’s Socket API [19,25]. The basic functions of the Socket API are
summarized in Table 1. These functions can be grouped into address handling
and protocol handling. In the first group, the function
i n t i n e t p t o n ( i n t f a m i l y , const char∗ s r c , void ∗ d s t )
converts the string src to the IP address dst. dst is addr in for IPv4, or addr in6
for IPv6; the type depends on the family parameter setting: AF INET (IPv4)
or AF INET6 (IPv6). The reverse direction is provided by
const char∗ i n e t n t o p ( i n t f a m i l y , const void ∗ s r c ,
char∗ dst , s o c k l e n t s i z e )
1010 Y. Luo et al.
Listing 1. The addrinfo Structure

struct a d d r i n f o {
int ai flags ;
int ai family ;
int ai socktype ;
int ai protocol ;
socklen t ai addrlen ;
struct s o c k a d d r ∗ ai addr ;
char∗ ai canonname ;
struct a d d r i n f o ∗ ai next ;
};
It converts the IP address src to the string dst, with a maximum length of
size. Note, both functions only handle addr in and addr in6 – not sockaddr in
and sockaddr in6. The application therefore has to assemble the latter structures
by itself! Similar to converting strings and addresses, it is also possible to resolve
DNS names and reverse-lookup DNS names for IP addresses. This is provided
by the functions gethostbyname() (DNS name resolution, like e.g. www.nntb.no)
and gethostbyaddr() (DNS reverse lookup). Originally, these functions have only
been defined for IPv4. Most operating systems, however, provide extended IPv6-
capable functions as well. This makes platform-specific code in multi-platform
applications necessary. Furthermore, gethostbyname() and gethostbyaddr() are
not thread-safe, creating further platform-specific variants. Therefore, the usage
of these functions is now deprecated. gethostbyname() and gethostbyaddr() have
been replaced by new functions, which also handle the address/string conversion
provided by inet pton() and inet ntop(), in a platform-independent way.
The getaddrinfo() function provides the translation of an address or DNS
name – as well as a port number or service name – into IP addresses, e.g.
sockaddr in (IPv4) or sockaddr in6 (IPv6) [19,25] as necessary:
i n t g e t a d d r i n f o ( const char∗ host ,
const char∗ serv ,
const struct a d d r i n f o ∗ hints ,
struct a d d r i n f o ∗∗ res )
That is, getaddrinfo() takes an address or DNS name string as host parameter,
a port number string or service name as serv parameter, an addrinfo structure
as hints parameter, as well as a pointer res to store the result (in form of
an addrinfo structure allocated by getaddrinfo()). The corresponding structure
addrinfo is shown in Listing 1. The hints structure can provide some specific
information for the address structure to be created, particularly the ai family
(AF INET or AF INET6 – but in most cases AF UNSPEC to support any IP
protocol). Furthermore two flags in ai flags can enforce the usage of IP address
only (AI NUMERICHOST, i.e. no DNS name allowed) and port number only
(AI NUMERICSERV, i.e. no service name allowed). We will later make this
clearer with examples. The resulting addrinfo structure contains the address in
Listing 2. An Example for getaddrinfo() with Address String and Port

const char∗ addressString = ” 2001:700:4100:2::179 ” ;
const char∗ portString = ” 443 ” ;
struct a d d r i n f o h i n t s ;
memset(& h i n t s , 0 , s i z e o f ( h i n t s ) ) ;
hints . ai family = AF UNSPEC ;
h i n t s . a i s o c k t y p e = SOCK STREAM;
hints . a i f l a g s = AI NUMERICHOST | AI NUMERICSERV ;
struct a d d r i n f o ∗ r e s u l t = NULL;
i f ( getaddrinfo ( addressString , portString ,
( const struct a d d r i n f o ∗)& h i n t s , &r e s u l t ) == 0 ) {
// s u c c e s s ; a d d r e s s i s i n r e s u l t −>a i a d d r !
freeaddrinfo ( result );
}
ai addr (i.e. sockaddr in for IPv4 or sockaddr in6 for IPv6) and its length in
ai addrlen. If there are multiple addresses (e.g. an IPv6 and an IPv4 address),
addrinfo structures may be chained using a pointer to the next addrinfo structure
in ai next. Since getaddrinfo() allocates storage space for the resulting addrinfo
structure(s), it is necessary to free them. This is realized by calling
void f r e e a d d r i n f o ( struct a d d r i n f o ∗ r e s )
The reverse direction, i.e. address to string or DNS name, is provided by:
int getnameinfo (const struct sockaddr∗ addr, socklen t addrlen,
char∗ host, socklen t hostlen ,
char∗ serv , socklen t servlen ,
int flags )
It takes a sockaddr in (IPv4) or sockaddr in6 (IPv6) structure in addr, its length
in addrlen and flags in flags. The result is written to buffers provided by
host for address string/hostname (with maximum length hostlen) and host for
port number string/service name (with maximum length servlen). Similar to
getaddrinfo(), the flag NI NUMERICHOST turns off DNS reverse lookup (the
result will be an address string); NI NUMERICSERV turns off service name
lookup (the result will be a port number string). We will make this clearer with
examples below.
The Socket API functions to create, bind and connect sockets, as well as
to write to and receive from addresses, are using sockaddr in (for IPv4) and
sockaddr in6 (for IPv6). Since these basics are part of various publications, like
e.g. [25], we do not go into details here. To write protocol-independent code,
it is therefore useful to let getaddrinfo() create all protocol-specific information
and let getnameinfo() make this information human-readable again. Ideally, the
application itself does not have to contain any IP-protocol-specific code.
In order to make the explanations above clearer, Listing 2 shows an example
to translate an IPv6 address string and a port number string to the corresponding
sockaddr in6 structure. getaddrinfo() obtains this information, in form of an
addrinfo structure (see Listing 1), pointing to the actual sockaddr in6 structure
in ai addr.
1012 Y. Luo et al.
Listing 3. An Example for getaddrinfo() with DNS Name and Service String
const char∗ nameString = ”www. nntb . no ” ;

const char∗ s e r v i c e S t r i n g = ” https ” ;
struct a d d r i n f o h i n t s ;
memset(& h i n t s , 0 , s i z e o f ( h i n t s ) ) ;
hints . ai family = AF UNSPEC ;
h i n t s . a i s o c k t y p e = SOCK STREAM;
struct a d d r i n f o ∗ r e s u l t = NULL;
i f ( g e t a d d r i n f o ( nameString , s e r v i c e S t r i n g ,
( const struct a d d r i n f o ∗)& h i n t s , &r e s u l t ) == 0 ) {
f o r ( const struct a d d r i n f o ∗ a i = r e s u l t ;
a i != NULL; a i = a i −>a i n e x t ) {
// I t e r a t e o v e r a l l a d d r e s s e s r e t u r n e d
}
freeaddrinfo ( result );
}
Listing 4. An Example for getnameinfo()

const struct s o c k a d d r ∗ a d d r e s s = r e s u l t −>a i a d d r ;
const s o c k l e n t a d d r e s s L e n g t h = r e s u l t −>a i a d d r l e n ;
char a d d r e s s S t r i n g [ NI MAXHOST ] ;
char p o r t S t r i n g [ NI MAXSERV ] ;
i f ( getnameinfo ( address , addressLength ,
( char∗)& a d d r e s s S t r i n g , s i z e o f ( a d d r e s s S t r i n g ) ,
( char∗)& p o r t S t r i n g , s i z e o f ( p o r t S t r i n g ) ,
NI NUMERICHOST | NI NUMERICSERV) == 0 ) {
p r i n t f ( ” Address : %s , Port : %s \n” , a d d r e s s S t r i n g , p o r t S t r i n g ) ;
}
Working with Internet applications, some servers are already using IPv6,
while some others still use IPv4. It is therefore convenient to use DNS names
instead of addresses. The example in Listing 3 therefore shows how to use
getaddrinfo() with DNS name (“www.nntb.no”) and service name (“https”)
instead. Calling getaddrinfo() is similar to Listing 2, except for not setting
the AI NUMERICHOST and AI NUMERICSERV flags (since we want to use
DNS and service lookup now). Since hosts may have multiple addresses in the
DNS, the example also iterates over all returned addresses chained by the
ai next pointer of the addrinfo structures. All sockaddr in and sockaddr in6
will contain the port number of service “https” (which is 443 – resolved from
/etc/services).
Finally, Listing 4 presents the usage of getnameinfo() to translate an address
(here: as result from getaddrinfo()) to a string, including the port number.
Obviously, not setting the flags NI NUMERICHOST and NI NUMERICSERV
would try to reverse-lookup the DNS name and convert the port number to a
service name.
3.2 Configuration for MPTCP

In order to use MPTCP on a Linux system, it is necessary to install an MPTCP-
capable kernel [29]. Currently, the mainline Linux kernel does not provide
MPTCP, so it is necessary to build a custom one based on the MPTCP devel-
opers’ patch1 .
In addition to using a kernel with MPTCP, it is furthermore necessary to
extend the routing configuration. Usually, all outgoing packets to an Internet
destination take a configured default route. This is, on a system being connected
to two different networks, all packets take the default route over one of the
networks (the one with the lowest metric). The other one is not used, unless the
primary one becomes unavailable. This is not desirable for MPTCP, which should
use both networks for load balancing with multi-path transport. It is therefore
necessary to configure multiple IP routing tables and select one of these tables
for a routed packet based on its source address.
To give an example, let a system have two IPv6 addresses A (in Network 1, via
interface NIC1) and B (in Network 2, via interface NIC2). Let both networks
be /64 networks for simplicity. Network 1 has router G1 as default gateway,
Network 2 has router G2 as default gateway. Therefore, two routing tables (1000
and 2000 for Network 1 and Network 2) are configured:
ip −6 r u l e add from $A t a b l e 1000
ip −6 r u l e add from $B t a b l e 2000
ip −6 r o u t e add $A/64 s c o p e l i n k dev $NIC1 table 1000
ip −6 r o u t e add d e f a u l t v i a $G1 dev $NIC1 table 1000
ip −6 r o u t e add $B/64 s c o p e l i n k dev $NIC2 table 2000
ip −6 r o u t e add d e f a u l t v i a $G2 dev $NIC2 table 2000
The first two lines configure the IP routing rules. If a packet is from source
address A, routing table 1000 is used. If it is from B, routing table 2000 is used.
The following lines configure local network and default route for both tables. Now,
MPTCP is able to send via both networks – by just using different source addresses.
4 Our Live Video Streaming System

For our live video system, we choose to use the Real-Time Messaging Proto-
col (RTMP) [23] over TCP. For this protocol, there is already a widely used
Open Source implementation available: SRS2 . Particularly, SRS supports the
structure of components we need:
• Push: the source of the live video. This could e.g. be a mobile user, filming
with his phone camera.
• Server: a server taking the video from the Push user, possibly adapting it
and distributing streams to viewers. In particular, this can be a cloud service
in an MEC setup, to provide low-latency processing, so that viewers get an
(almost) live stream.
• Client: a viewer, receiving the media stream from the server.
1
MPTCP Git repository: https://github.com/multipath-tcp/mptcp.
2
SRS Git repository: https://github.com/ossrs.
1014 Y. Luo et al.
However, SRS was solely based on TCP over IPv4, i.e. without support for
IPv6. As part of the work for this paper, we added IPv6 support as well. The
changes for this support have been submitted as pull request3 to the upstream
project. It has already been merged into the upstream SRS sources, i.e. IPv6
support is now available for all SRS users. Basically, the idea of the changes
for SRS was to replace all IPv4-specific code by IP-protocol-independent code
based on getaddrinfo() and getnameinfo() (see Sect. 3). That is, getaddrinfo() is
now used to convert address strings (see Listing 2) or resolve host names (see
Listing 3) to either sockaddr in (IPv4) or sockaddr in6 (IPv6). Then, the appli-
cation code can create sockets of the right family (AF INET or AF INET6 –
returned by getaddrinfo() in ai family of the addrinfo structure; see Listing 1).
In the reverse direction, socket addresses are translated into strings by getname-
info() (see Lising 4). By making the code IP-protocol-independent, we added
IPv6 support. This not only simplified the original code, but also made it more
portable as well.
We furthermore adapted bwm-ng4 , a tool to record the data statistics of
network interfaces, with an improved accuracy of the recorded information. This
change had also been provided as pull request5 and been merged by the upstream
project.
Finally, for the video display at the client, we use ffplay. It is a playback
component of the well-known FFmpeg6 tools.
5 Proof of Concept Evaluation

To evaluate the live broadcast system with SRS over MPTCP, we have set up
the scenario as shown in Fig. 1: it consists of RTMP Push, Server and Client.
In the Push-Server network, as well as in the Server-Client network, each device
is equipped with two network interface cards (NIC). Each NIC can be config-
ured with an IPv4 or IPv6 address. Between Push and Server, as well as between
Server and Client, MPTCP can therefore establish 4 subflows. We consider NICs
corresponding to different Internet service providers (ISP), i.e. all NIC1s corre-
spond to ISP1 and all NIC2s correspond to ISP2. Then, the setup corresponds
to a 2-ISP mobile edge setup, with the Push device e.g. a user sending a live
video stream, the Server device a media server in a (mobile edge) computing
centre and the Client device another user watching the live video. The video is
observed by a human user, to check for disruptions and quality problems. To
assess the performance of our setup, we apply the testing sequence as shown in
Fig. 2. It consists of 4 phases: (1) load balancing with all interfaces up; (2) ISP2
is down (all traffic for NIC2 of all devices is blocked); (3) ISP1 goes down, at
the same time ISP2 comes up (all traffic of the NIC1s blocked); (4) load balanc-
ing with all interfaces up again. With these phases, we want to mimic simple
3
IPv6 support for SRS: https://github.com/ossrs/srs/pull/988 (pull request).
4
bwm-ng Git repository: https://github.com/vgropp/bwm-ng.
5
CSV file output improvement: https://github.com/vgropp/bwm-ng/pull/20 (pull
request).
6
FFmpeg: https://www.ffmpeg.org.
Fig. 1. The structure of the streaming platform
NIC
Seamless Switching
Load Balancing
NIC1
Robustness
NIC2
Time
1 2 3 4
Fig. 2. Testing scenario sequence diagram
handover and multi-ISP load balancing in a multi-homed mobile edge scenario.

In the following measurement, each phase lasts 25 s. For each NIC, the outgoing
and incoming amount of data is recorded by using bwm-ng. On the Client device,
ffplay is used to display the streamed video.
Figure 3 presents the IPv6 results for the Push side (outgoing data rate) and
Client side (incoming data rate). In the first 25 s, load balancing is performed
by MPTCP. As intended, MPTCP distributes the traffic to both networks. The
outage of ISPs at t = 25 s (by blocking all NIC1s) is properly handled by using
the other network. The data rate there increases, since the video quality remains
unchanged. At t = 50 s, the ISPs change (ISP1 down, ISP2 up again). Particularly
at the Client side, it is shown that it takes a moment for the system to detect
the working ISP2 network and to move all traffic to the NIC2s. Once MPTCP
detects that the NIC2s can be used, instead of the NIC1s, the RTMP system can
quickly fill up the client’s playback buffer (see the data rate peak). The video
client at the Client device uses a buffer of 64 MiB, i.e. this playback buffer is able
to cover the time needed by MPTCP for changing the networks. The human user
did not notice a disruption of his video. At t = 75 s, the blocking of the NIC1s
is ended. The system went back to load balancing as in the first phase.
In summary, we can show that (1) MPTCP+IPv6 can provide load balancing
in multi-ISP real-time streaming setups; (2) MPTCP+IPv6 can support seamless
handover in failure scenarios; (3) There is good compatibility among IPv6 in the
Network Layer, MPTCP in the Transport Layer and RTMP in the Application
Layer. The configuration therefore seems to be a reasonable choice for live real-
time streaming with mobile edge computing.
1016 Y. Luo et al.
1000 1250 Interface I
I=NIC 1
I=NIC 2
Outging Data Rate D [KiB/s]
750
500
250
0
0 5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90 95 100
Relative Time t [s]
(a) Push Side

Interface I
1000 1250
I=NIC 1
Incoming Data Rate D [KiB/s]
I=NIC 2
750
500
250
0
0 5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90 95 100
Relative Time t [s]
(b) Client Side
Fig. 3. Outgoing per-interface data rate with IPv6

This paper introduced our live broadcast platform based on RTMP, MPTCP and
IPv6. As a proof of concept, we showed that the system is able to handle typical
load balancing, robustness and handover situations of multi-ISP scenarios.
As ongoing work on video transport over MPTCP with IPv6, we are now
going to test and evaluate our setup in realistic, larger-scale Internet setups by
using the NorNet Core testbed [6,7,9,21]. The NorNet Core testbed con-
sists of test nodes being located at different sites in multiple countries. Test appli-
cations can experience the same quality of service (QoS) as “normal” Internet
users. The NorNet Core setup allows tests under realistic Internet conditions.
We particularly plan to also combine it with MELODIC (Multi-Cloud Execution-
Ware for Large-scale Optimised Data-Intensive Computing) [10,13], to support
the cloud infrastructure for mobile edge computing. MELODIC is a middleware
for application deployment in multi-cloud setups. Our idea is therefore to use
MELODIC to provide a scalable Server-side setup in the MEC cloud (e.g. for
video transcoding, caching and editing). The resources can then be instantiated
on demand and multi-cloud support provides cost-effective deployment without
vendor lock-in.
Furthermore, we want to apply NEAT (New, Evolutive API and Transport-

Layer architecture for the Internet) technology [8,11,28]. The goal is to allow
network and transport “services” provided to applications – such as reliabil-
ity, low-delay communication or security – to be dynamically tailored based on
application demands, current network conditions, hardware capabilities or local
policies. For example, NEAT may choose to use advanced technologies like multi-
path transport with MPTCP and IPv6 where available, but fall back to regular
TCP in legacy networks.
Finally, with the application of Reliable Server Pooling (RSerPool) [4,12,14], it
would also be possible to achieve high-availability by redundant Server-instances.
RSerPool could then also be used to balance sessions between several Server-
instances.
References
1. Becke, M., Dreibholz, T., Adhari, H., Rathgeb, E.P.: On the fairness of transport
protocols in a multi-path environment. In: Proceedings of the IEEE International
Conference on Communications (ICC), Ottawa, Ontario, Canada, pp. 2666–2672
(2012)
2. Chen, J.L., Chao, H.C., Kuo, S.Y.: IPv6: more than protocol for next generation
internet. Comput. Commun. 29, 3011–3012 (2006)
3. Deering, S.E., Hinden, R.M.: Internet Protocol, Version 6 (IPv6) Specification.
Standards Track RFC 2460, IETF (1998)
4. Dreibholz, T.: Reliable server pooling – evaluation, optimization and extension of
a novel IETF architecture. Ph.D. thesis, University of Duisburg-Essen, Faculty
of Economics, Institute for Computer Science and Business Information Systems
(2007)
5. Dreibholz, T.: Evaluation and optimisation of multi-path transport using the
stream control transmission protocol. Habilitation treatise, University of Duisburg-
Essen, Faculty of Economics, Institute for Computer Science and Business Infor-
mation Systems (2012)
6. Dreibholz, T.: NorNet – building an inter-continental internet testbed based on
open source software. In: Proceedings of the LinuxCon Europe, Berlin, Germany
(2016)
7. Dreibholz, T.: NorNet – the internet testbed for multi-homed systems. In: Pro-
ceedings of the Multi-Service Networks Conference (MSN, Coseners). Abingdon,
Oxfordshire (2016)
8. Dreibholz, T.: A practical introduction to NEAT at Hainan University. Invited
Talk at Hainan University, College of Information Science and Technology (CIST)
(2017)
9. Dreibholz, T.: An introduction to multi-path transport at Hainan University.
Keynote Talk at Hainan University, College of Information Science and Technology
(CIST) (2017)
10. Dreibholz, T.: Big data applications on multi-clouds: an introduction to the
MELODIC project. Keynote Talk at Hainan University, College of Information
Science and Technology (CIST) (2017)
11. Dreibholz, T.: NEAT sockets API. Internet Draft draft-dreibholz-taps-neat-
socketapi-04, IETF, Individual Submission (2019)
1018 Y. Luo et al.
12. Dreibholz, T., Becke, M.: The RSPLIB project – from research to application. In:
Demo Presentation at the IEEE Global Communications Conference (GLOBE-
COM) (2010)
13. Dreibholz, T., Mazumdar, S., Zahid, F., Taherkordi, A., Gran, E.G.: Mobile edge as
part of the multi-cloud ecosystem: a performance study. In: Proceedings of the 27th
Euromicro International Conference on Parallel, Distributed and Network-Based
Processing (PDP), Pavia, Lombardia, Italy (2019)
14. Dreibholz, T., Zhou, X., Becke, M., Pulinthanath, J., Rathgeb, E.P., Du, W.: On
the security of reliable server pooling systems. Int. J. Intell. Inf. Database Syst.
(IJIIDS) 4(6), 552–578 (2010)
15. Dreibholz, T., Zhou, X., Fu, F.: Multi-path TCP in real-world setups – an eval-
uation in the NorNet core testbed. In: 5th International Workshop on Protocols
and Applications with Multi-Homing Support (PAMS), Gwangju, South Korea,
pp. 617–622 (2015)
16. Ford, A., Raiciu, C., Handley, M., Bonaventure, O.: TCP extensions for multipath
operation with multiple addresses. RFC 6824, IETF (2013)
17. Fu, F., Zhou, X., Dreibholz, T., Wang, K., Zhou, F., Gan, Q.: Performance compar-
ison of congestion control strategies for multi-path TCP in the NorNet testbed. In:
Proceedings of the 4th IEEE/CIC International Conference on Communications
in China (ICCC), Shenzhen, Guangdong, People’s Republic of China, pp. 607–612
(2015)
18. Fu, F., Zhou, X., Tan, Y., Dreibholz, T., Adhari, H., Rathgeb, E.P.: Performance
analysis of MPTCP protocol in multiple scenarios. Comput. Eng. Appl. 52(5),
89–93 (2016)
19. Gilligan, R., Thomson, S., Bound, J., McCann, J., Stevens, W.R.: Basic socket
interface extensions for IPv6. Informational RFC 3493, IETF (2003)
20. Golkar, F., Dreibholz, T., Kvalbein, A.: Measuring and comparing internet path
stability in IPv4 and IPv6. In: Proceedings of the 5th IEEE International Confer-
ence on the Network of the Future (NoF), Paris, France, pp. 1–5 (2014)
21. Gran, E.G., Dreibholz, T., Kvalbein, A.: NorNet core - a multi-homed research
testbed. Comput. Netw. Spec. Issue Future Internet Testbeds 61, 75–87 (2014)
22. Livadariu, I.A., Ferlin, S., Alay, Ö., Dreibholz, T., Dhamdhere, A., Elmokashfi,
A.M.: Leveraging the IPv4/IPv6 identity duality by using multi-path transport.
In: Proceedings of the 18th IEEE Global Internet Symposium (GI) at the 34th
IEEE Conference on Computer Communications (INFOCOM), Hong Kong, Peo-
ple’s Republic of China, pp. 312–317 (2015)
23. Parmar, H., Thornburgh, M.: Adobe’s real time messaging protocol. Technical
report, Adobe (2012)
24. Raiciu, C., Paasch, C., Barré, S., Ford, A., Honda, M., Duchêne, F., Bonaventure,
O., Handley, M.: How hard can it be? Designing and implementing a deployable
multipath TCP. In: Proceedings of the 9th USENIX Conference on Networked
Systems Design and Implementation (NSDI), San Jose, California, USA, pp. 1–14
(2012)
25. Stevens, W.R., Fenner, B., Rudoff, A.M.: Unix Network Programming. Addison-
Wesley Professional, Boston (2003)
26. Tadayoni, R., Henten, A.: From IPv4 to IPv6: lost in translation? Telemat. Inf.
33(2), 650–659 (2016)
27. Wang, K., Dreibholz, T., Zhou, X., Fu, F., Tan, Y., Cheng, X., Tan, Q.: On the path
management of multi-path TCP in internet scenarios based on the NorNet testbed.
In: Proceedings of the IEEE International Conference on Advanced Information
Networking and Applications (AINA), Taipei, Taiwan, People’s Republic of China,
pp. 1–8 (2017)
28. Weinrank, F., Grinnemo, K.J., Bozakov, Z., Brunström, A., Dreibholz, T., Hurtig,
P., Khademi, N., Tüxen, M.: A NEAT way to browse the web. In: Proceedings
of the ACM, IRTF and ISOC Applied Networking Research Workshop (ANRW),
Praha, Czech Republic, pp. 33–34 (2017)
29. Wischik, D., Raiciu, C., Greenhalgh, A., Handley, M.: Design, implementation and
evaluation of congestion control for multipath TCP. In: Proceedings of the 8th
USENIX Conference on Networked Systems Design and Implementation (NSDI),
Boston, Massachusetts, USA, pp. 99–112 (2011)
30. Zhou, F., Dreibholz, T., Zhou, X., Fu, F., Tan, Y., Gan, Q.: The performance
impact of buffer sizes for multi-path TCP in internet setups. In: Proceedings of the
IEEE International Conference on Advanced Information Networking and Appli-
cations (AINA), Taipei, Taiwan, People’s Republic of China, pp. 9–16 (2017)
Towards Realistic Simulations
of Arbitrary Cross-Cloud Workloads
Nicolay Mohebi1 and Feroz Zahid2(B)

1
USIT, University of Oslo,
Kristen Nygaards hus, Gaustadalléen 23a, 0373 Oslo, Norway
nicolay.mohebi@usit.uio.no
2
Simula Research Laboratory, Martin Linges vei 25, 1364 Fornebu, Norway
feroz@simula.no
Abstract. The research undertakings in cloud computing often require

designing new algorithms, techniques, and solutions requiring large-scale
cloud deployments for comprehensive evaluation. Simulations make a
powerful and cost-effective tool for testing, evaluation, and repeated
experimentation for new cloud algorithms. Unfortunately, even though
cloud federation and hybrid cloud simulations are explored in the litera-
ture, Cross-Cloud simulations are still largely an unsupported feature in
most popular cloud simulation frameworks.
In this paper, we present a Cross-Cloud simulation framework, which
makes it possible to test scheduling and reasoning algorithms on Cross-
Cloud deployments with arbitrary workload. The support of Cross-Cloud
simulations, where individual application components are allowed to be
deployed on different cloud platforms, can be a valuable asset in select-
ing appropriate mixture of cloud services for the applications. We also
implement a Cross-Cloud aware reasoner using our Cross-Cloud simu-
lation framework. Simulations using both simple applications and com-
plex multi-stage workflows show that the Cross-Cloud aware reasoner
can substantially save cloud usage costs for most multi-component cloud
applications.
1 Introduction
The idea that computational power could be provided as a utility to the users
dates back to 1966 [1]. However, it was not until recently that the cloud comput-
ing systems started gaining popularity, thanks mainly to the better virtualization
support and the rise of the Internet. Over the last few years, cloud computing
has fundamentally brought a paradigm shift in computing, and the industry has
witnessed an accelerated transition from traditional cluster computing architec-
tures to highly flexible cloud-based services.
Clouds bring significant advantages over traditional cluster computing archi-
tectures including flexibility, ease of setup and deployment, high-availability, and
on-demand resource allocation [2]. In addition, a highly attractive pay-as-you-
go [3] business model of the cloud computing drives a large number of enter-
prises to migrate to cloud computing solutions. However, as it is true for any
https://doi.org/10.1007/978-3-030-15035-8_100
Towards Realistic Simulations of Arbitrary Cross-Cloud Workloads 1021
other emerging technology, cloud computing is full of interesting research chal-

lenges [4]. Novel algorithms, techniques, and solutions are required to further
an even broader cloud adoption. For instance, energy efficiency, workload opti-
mization, performance predictability, security, and resource utilization, among
others, are some areas where improvements are needed. As clouds work on a large
scale with hundreds of thousands of virtual machines (VMs) a common place,
research undertakings requiring repeated experimentation on physical hardware
often result in prohibitively high costs. Simulations make a viable alternative
opening doors for cost-effective evaluation of new research in the domain of
cloud computing. Hence, several good cloud simulation frameworks have been
developed over the years for modeling and simulation of workloads running on
cloud computing data centers [5,6].
With the emerging cloud options, modern enterprises increasingly rely on
hybrid cloud solutions to meet their computational demands by acquiring addi-
tional resources from public clouds dynamically as per their needs. International
Data Corporation (IDC), which is a leading market-research firm, in its Cloud-
View Survey 2017 [7], reported that 87% of cloud users have adapted a hybrid
cloud strategy and 56% of the users use more than one type of cloud deployment.
In general, cloud federation enables end users to integrate resources from differ-
ent cloud systems. Federated clouds offer more freedom to the cloud users and
increase the granularity of choices in the application deployment. With Cross-
Cloud1 federated application deployments, cloud users no longer need to live
with a single cloud provider, and accept whatever pricing schemes or Service
Level Agreement (SLAs) are imposed, but could choose the best match for each
individual job, or even dynamically spread a single application across multiple
clouds to get the job done in the most efficient, secure and cost-effective manner.
From the cloud user’s perspective, support of generalized Cross-Cloud simu-
lations can be valuable to select appropriate cloud services for the applications.
Simulating Cross-Cloud workloads, however, is either not available or not well-
supported on available cloud simulation frameworks [4,8,9]. Apart from a few
use-case specific extensions to the CloudSim [5], which is a popular cloud sim-
ulation framework, the support for Multi-Cloud and Cross-Cloud simulations
is not readily available. The CloudSim extensions [10–12] too, as discussed in
Sect. 2.2, are very specific to the research use-cases and easily implementing new
Cross-Cloud simulations is not possible using the available code-base. Moreover,
CloudSim itself has been forked out numerous times from the main project, and
segregated features are implemented in separate research projects with each fork
maintaining its own copy, making it harder to re-use.
1
We use the term Cross-Cloud to refer to application deployments where multiple
cloud platforms are simultaneously used to deploy application components. The
term Multi-Cloud is also popular, though. We differentiate Multi-Cloud scenarios
from Cross-Clouds – in Multi-Clouds, applications are capable of being deployed on
different cloud platforms, but one at a time, contrary to the Cross-Cloud deployments
of application components on segregated cloud platforms at the same time.
1022 N. Mohebi and F. Zahid
In this paper, we present a framework for Cross-Cloud simulations based

on CloudSimPlus [6]. The CloudSimPlus simulation framework is a new re-
engineered version of CloudSim, implementing software engineering standards
and design principles to ensure code reusability, maintainability, and extensi-
bility. Further, based on our implemented Cross-Cloud simulation framework,
we develop a Cross-Cloud deployment reasoner. The job of the reasoner is to
ensure that the most cost-effective resources are selected from available cloud
service providers (CSPs) for each application component, depending on its com-
putational requirements, prices of the available VM options at different CSPs,
and taking into consideration its communication affinities with other application
components. Simulations using both simple applications and complex multi-stage
workflows show that the Cross-Cloud aware reasoner can substantially save cloud
usage costs.
The rest of this paper is structured as follows. In Sect. 2 we provide a brief
background and the related work. We present the design of our Cross-Cloud
simulation framework and its implementation details in Sect. 3. In Sect. 4, we
provide a comprehensive evaluation of the Cross-Cloud-aware reasoning in our
Cross-Cloud simulation framework. We conclude in Sect. 5.

2.1 Cloud, Multi-cloud, and Cross-Cloud
Broadly speaking, clouds come in three flavors: public, private, and hybrid. The
public clouds, owned by the CSPs, offer infrastructure and services to their
customers over the Internet. The basic advantage of using a public cloud, as
mentioned above, is that the organizations need not to invest a large capital
expenditure to buy and setup the hardware needed to run their applications. On
the other hand, the private clouds are owned and operated by individual com-
panies and organizations. In private clouds, cloud orchestration is implemented
over private hardware resources to allow for efficient utilization of the available
hardware among applications owned and operated by the company. A hybrid
cloud is a combination of private and public clouds in which resources acquired
from public clouds are used to complement the available hardware in the private
infrastructure. For example, hybrid cloud setups can dynamically utilize public
clouds for application cloud bursting [13] in high-load situations. Applications
capable of deploying on Multi-Clouds use an abstract interface to the application
programming interfaces (APIs) of the specific CSPs so they can easily migrate
to any cloud platform. Finally, Cross-Cloud deployments refer to the application
deployments spanning multiple clouds, where each individual application com-
ponent can be deployed independently of the deployment of other application
components.
2.2 Related Work

Cloud simulations is an interesting topic in the literature and a large number of
cloud simulation frameworks have been developed over the years, mainly based
on discrete event simulations. Popular cloud simulation frameworks include

CloudSim [5], GreenCloud [14], iCanCloud [15], among others. The CloudSim, in
particular, has gained significant popularity among cloud researchers. As a result
of its unique features and position, many researchers used and extended it for
their research purposes. CloudSim has been extended to include support for fed-
erated cloud resources [16], workflow preparation and execution [17], edge [18],
among others. Many of the research projects based on CloudSim eventually
forked out from the main project and the corresponding contributors kept main-
taining the forked copies without merging back into the main project, resulting
in code duplications and different code styling among various projects. To cater
this, recently a new re-engineered version of CloudSim v3, CloudSimPlus [6], has
been developed. CloudSimPlus sets up design principles based on software engi-
neering best practices to ensure that the CloudSimPlus code remains reusable,
maintainable, and extensible.
Cross-Cloud simulations have limited support in existing cloud simulation
frameworks. In [10], Buyya et al. proposed an architectural framework, based
on CloudSim, that supports cost-aware VM provisioning on Multi-Clouds, but
the implementation focuses mainly on software-defined networking and network
function virtualization, and generalized workload modeling and execution is not
available. Another architecture for workload simulations in Cross-Clouds envi-
ronments was proposed by Grozev et al. in [11], but it only supports simulation
of three-tier applications, limiting its general-purpose usage. Similarly, CEP-
Sim [12] supports execution of complex event processing queries on different
cloud environments including Multi-Clouds and Cross-Cloud under diverse load
conditions. However, that too does not support generating generalized workload
and Cross-Cloud simulation.
3 Cross-Cloud Simulation Framework

Our Cross-Cloud simulation framework, CrossCloudSim-Plus, is an extension of
CloudSim-Plus that aims to enable support of Cross-Cloud simulations with a
focus on generalized workload execution, and pluggable architecture for imple-
menting resource reasoners. CrossCloudSim-Plus adheres to the same design
principles as CloudSim-Plus, making it easy to maintain, re-use, and extend as
per the user needs. CloudSim-Plus lets you model cloud infrastructures by defin-
ing DataCenter s and Hosts, but lacks the notion of Clouds as their own enti-
ties in the simulation, a feature necessary for Cross-Cloud simulations. More-
over, workload in CloudSim-Plus only consists of single cloudlets, and hence
advanced applications consisting of multiple components cannot be easily mod-
eled. CrossCloudSim-Plus introduces both clouds as first class simulation entities
and complex applications as possible workloads. The new classes and interfaces
introduced in CrossCloudSim-Plus do not modify implementations below the
DataCenter layer, ensuring the execution results are comparable and verifiable
using the original CloudSim-Plus.
1 {
2 "name": "p2.xlarge",
3 "cpus": 4,
4 "memoryInMB":62464,
5 "pricePerHour":0.900,
6 "csp": "AmazonEC2",
7 "region": "US-East(Ohio)"
8 }
Listing 1: An example instance definition in json.
3.1 Cloud Service Providers
In CloudSim-Plus, customers are represented through CloudDatacenterBrokers

that handle submission of cloudlets to the different data centers according to
an allocation policy. CrossCloudSim-Plus introduces the CloudServiceProvider-
Broker, a broker that handles submission between the different CloudServi-
ceProviders. A Cloud Service Provider is modeled through the CloudService-
Provider class. It consists of a collection of data centers, and a set of CloudDat-
acenterBrokers, one for each of the CloudServiceProviderBroker s.
3.2 Workloads
Workload in CrossCloudSim-Plus may consist of single cloudlets, applications,

and workflows. Cloudlets consist of a set of requirements and execution prop-
erties, and may be assigned to an instance type. An InstanceType represents a
VM instance that the cloudlets are executed on. The VM instance templates
are defined using a JSON-based file. In our framework implementation, we pro-
vided VM templates for all current VM offerings for Amazon AWS and Microsoft
Azure. However, new VM templates can easily be added. An example VM tem-
plate is given in Listing 1.
Application represents an application in the form of a graph. The vertices
in the graph represent application components executing one or more cloudlets,
while an edge represents the communication link between two application com-
ponents. The link has a cost tied to it representing the communication cost on
between two components of the application. This cost is a value in the unit
interval that affects the runtime of the cloudlets in both vertices, which in turn
affects an application’s total runtime. Workflows are modeled as directed acyclic
graphs. The vertices may contain one or more cloudlets, and the edges represent
the transfer cost related to transferring the results from the source vertex to the
target vertex. The data transfer time is also added to the cloudlet’s runtime.
The cloudlets are submitted dynamically during the execution of the simulation
as the required jobs finish.
Fig. 1. An overview of the Cross-CloudSimplus framework
3.3 Reasoning Framework
In addition to clouds and workload improvements, we introduce a framework

for enabling reasoning on sets of instance types and workloads. Allowing for
implementation of algorithms to assign cloudlets to a corresponding instance
type based on different criterias, e.g. execution time, cost. The instance types
used are taken from real-world cloud service providers. While reasoning on single
cloudlets are pretty straightforward, applications consisting of multiple edges
need to take into account the communication costs associated with them.
We implemented a Cross-Cloud Reasoner that takes VM instance prices and
the communication links represented in the application graph into account to
find out the best possible resource allocations for a particular workload.
An overview of the CrossCloudSim-Plus framework with marked changes and
extensions to the CloudSim-Plus is shown in Fig. 1.
4 Evaluation
In order to evaluate CrossCloudSim-Plus, we implement a Cross-Cloud aware

reasoner and ran simulation experiments for different sets of applications and
workflows. We define different pricing for VMs in all data centers for Amazon
AWS and Microsoft Azure, based on available pricing information. The basic
parameters defining each simulation experiment are kept constant while cloudlet
sizes are randomly chosen from a range.
4.1 Cross-Cloud Applications
We define cloud applications as a graph, G(V, E), where the set of vertices, V ,
describes self-containing application components that can be deployed on a VM
in a cloud environment. The set of edges between the applications components,
E, refers to the communication links between corresponding application compo-
nents. For the sake of simplicity, we assume that all the communication links are
bi-directional. One or more cloudlets are then executed at each application com-
ponent. Using the functionality provided in our Cross-Cloud simulation frame-
work, each of the application components can be deployed independently on a
suitable VM on any of the available cloud data centers of the CSPs. However, if
only the cheapest VMs are chosen satisfying cloudlet demands, the overhead of
the communication between geographically-dispersed data centers may outrun
the cost benefits achieved by choosing the cheapest VM options.
Mean Cost (USD)
Mean Cost (USD)

80
400
Small Apps (Simple Reasoner) Large Apps (Simple Reasoner)

Small Apps (Cross−Cloud Reasoner) Large Apps (Cross−Cloud Reasoner)
Medium Apps (Simple Reasoner) Very Large Apps (Simple Reasoner)
60
Medium Apps (Cross−Cloud Reasoner) Very Large Apps (Cross−Cloud Reasoner)

300
40
200
100
20
0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
Component Connectivity Component Connectivity
(a) Small and Medium sized applications (b) Large sized applications
Fig. 2. Cross-Cloud-aware deployment reasoning saves substantial cloud costs.
To demonstrate cost-efficiency of Cross-Cloud Reasoner over Simple Reasoner

which selects the cheapest instance types satisfying demands, we simulated four
different types of applications based on their size. The small-sized applications
have a maximum of 5 application components and up to 10 cloudlets per appli-
cation. The medium and large applications have maximum 10 and 15 application
components, and up to 20 and 30 cloudlets, respectively. The very large appli-
cations have up to 20 components and 40 cloudlets. The exact number of appli-
cation components, number of cloudlets, and the communication links between
application components, however, are chosen randomly in each of the simulated
experiments. Experiments are run 100 times and average results are presented.
The average costs of the simulated experiments for small- and medium-sized
applications is presented in Fig. 2(a). The x -axis shows component connectivity
which refers to the percentage of the communication links implemented in the
simulation experiments. As shown in the figure, for the small-sized applications,
the impact of the Cross-Cloud reasoning is limited (shown as squared solid line
for Cross-Cloud Reasoner and dashed circled line for the Simple cost-based Rea-
soner). However, for the medium-sized application, as shown with crossed solid
and dash line with a plus sign for Cross-Cloud and Simple Reasoner respectively,
the impact is huge. In particular, as the connectivity between the application
components increases, the cost of the applications increases very rapidly for the
Simple Reasoner. The Cross-Cloud Reasoner on the other hand is minimally
affected by the communication links, as it considers the application connectivity
into account while provisioning the VMs for the applications. The results for
the large and very large applications are shown in Fig. 2(b). Here too, we see
that the cost of the applications when Simple cost-based reasoning is used is
higher as compared to the Cross-Cloud reasoning. The impact is greater for the
large-sized applications and with higher percentage of connectivity.
4.2 Multi-stage Workflows
We also implement multi-stage workflows where different set of application com-

ponents are connected to each other in different application stages. To elaborate
further, each workflow stage consists of a different set of connectivity graph
between application components, and the stages are executed in the cloud as
per a given order. Multi-stage work flows are useful to show the impact of the
simple cost-based reasoning on the e-science applications, which often consist of
dependent application stages.
Mean Cost (USD)
Mean Cost (USD)
Simple Reasoner Simple Reasoner

250
250
Cross−Cloud Reasoner Cross−Cloud Reasoner

150
150
0 50
0 50
2 3 4 5 2 3 4 5
Workflow Depth Workflow Depth
(a) Small-sized workflows (b) Medium-sized workflows

Mean Cost (USD)
Mean Cost (USD)
Simple Reasoner Simple Reasoner

250
250
Cross−Cloud Reasoner Cross−Cloud Reasoner

150
150
0 50
0 50
2 3 4 5 2 3 4 5
Workflow Depth Workflow Depth
(c) Large-sized workflows (d) Very large-sized workflows
Fig. 3. Cross-Cloud workflows benefit from workflow-aware scheduling.

As described in Sect. 4.1, we used different application sizes for the workflow-
based simulations as well. The work-flow application sizes are based on the max-
imum number of application components in each of the applications. A number
of the application components are then connected in each stage. The workflow
depth defines the number of layers of workflow graphs that are defined for an
application. We present the results of the workflow-aware scheduling in Fig. 3.
The small sized applications, as shown in Fig. 3(a) are the least affected. How-
ever, still, with a workflow depth of 5, the Cross-Cloud reasoner costs about $51,
which is about 28% improvement over $71 costs of the simple reasoning. The
results for medium- and large- sized workflows are given in Fig. 3(b) and (c),
respectively. The very large applications, as shown in Fig. 3(d), benefits hugely
with the help of Cross-Cloud reasoner. As presented in the figure, the 5-layer
very large workflow on average in our simulated experiments cost around $131,
which is about 1.7 times better than the cost using the simple reasoner (which
is about $301).
5 Conclusion
In this paper, we tackled the challenge of designing a Cross-Cloud simulation
framework, CrossCloudSim-Plus, where application components are allowed to
be deployed on multiple cloud platforms. In the future, we plan to further extend
our simulation framework and include support for more complex application
design through modeling, and extend our reasoning framework to include data-
aware reasoning.
Acknowledgement. This work has received funding from the European Union’s
H2020 programme under grant agreement no. 731664 (MELODIC).
References
1. Parkhill, D.F.: The Challenge of the Computer Utility (1966)
2. Mell, P., Grance, T.: The NIST definition of cloud computing. Nat. Inst. Stand.
Technol. 53(6), 50 (2009)
3. Rappa, M.A.: The utility business model and the future of computing services.
IBM Syst. J. 43(1), 32 (2004)
4. Taherkordi, A., Zahid, F., Verginadis, Y., Horn, G.: Future cloud systems design:
challenges and research directions. IEEE Access (2018). https://doi.org/10.1109/
ACCESS.2018.2883149
5. Calheiros, R.N., Ranjan, R., Beloglazov, A., De Rose, C.A., Buyya, R.: CloudSim:
a toolkit for modeling and simulation of cloud computing environments and evalu-
ation of resource provisioning algorithms. Softw. Pract. Exp. 41(1), 23–50 (2011)
6. Silva Filho, M.C., Oliveira, R.L., Monteiro, C.C., Inácio, P.R., Freire, M.M.:
CloudSim plus: a cloud computing simulation framework pursuing software engi-
neering principles for improved modularity, extensibility and correctness. In: 2017
IFIP/IEEE Symposium on Integrated Network and Service Management (IM), pp.
400–406. IEEE (2017)
7. IDC: CloudView Survey 2017. https://www.idc.com/getdoc.jsp?containerId=

prUS42878417
8. Fakhfakh, F., Kacem, H.H., Kacem, A.H.: Simulation tools for cloud computing: a
survey and comparative study. In: IEEE/ACIS 16th International Conference on
Computer and Information Science (ICIS), pp. 221–226. IEEE (2017)
9. Zhao, W., Peng, Y., Xie, F., Dai, Z.: Modeling and simulation of cloud computing:
a review. In: IEEE Asia Pacific Cloud Computing Congress (APCloudCC), pp.
20–24. IEEE (2012)
10. Buyya, R., Son, J.: Software-defined multi-cloud computing: a vision, architectural
elements, and future directions. arXiv e-prints (2018)
11. Grozev, N., Buyya, R.: Performance modelling and simulation of three-tier appli-
cations in cloud and multi-cloud environments. Comput. J. 58 (2013). https://doi.
org/10.1093/comjnl/bxt107
12. Higashino, W.A., Capretz, M.A., Bittencourt, L.F.: CEPSim: modelling and sim-
ulation of complex event processing systems in cloud environments. Future Gener.
Comput. Syst. 65, 122–139 (2016)
13. Guo, T., Sharma, U., Shenoy, P., Wood, T., Sahu, S.: Cost-aware cloud bursting
for enterprise applications. ACM Trans. Internet Technol. (TOIT) 13(3), 10 (2014)
14. Kliazovich, D., Bouvry, P., Khan, S.U.: GreenCloud: a packet-level simulator of
energy-aware cloud computing data centers. J. Supercomput. 62(3), 1263–1283
(2012)
15. Núñez, A., Vázquez-Poletti, J.L., Caminero, A.C., Castañé, G.G., Carretero, J.,
Llorente, I.M.: iCanCloud: a flexible and scalable cloud infrastructure simulator.
J. Grid Comput. 10(1), 185–209 (2012)
16. Kohne, A., Spohr, M., Nagel, L., Spinczyk, O.: FederatedCloudSim: a SLA-aware
federated cloud simulation framework. In: Proceedings of the 2nd International
Workshop on CrossCloud Systems, p. 3. ACM (2014)
17. Chen, W., Deelman, E.: Workflowsim: a toolkit for simulating scientific work-
flows in distributed environments. In: 8th International Conference on E-science
(e-Science), pp. 1–8. IEEE (2012)
18. Sonmez, C., Ozgovde, A., Ersoy, C.: EdgeCloudSim: an environment for perfor-
mance evaluation of Edge Computing systems. In: Second International Conference
on Fog and Mobile Edge Computing (FMEC), pp. 39–44. IEEE (2017)
Data Center Clustering
for Geographically Distributed
Cloud Deployments
Dipesh Pradhan and Feroz Zahid(B)
Simula Research Laboratory, Martin Linges vei 25, 1364 Fornebu, Norway
{dipesh,feroz}@simula.no
Abstract. Geographically-distributed application deployments are crit-

ical for a variety of cloud applications, such as those employed in the
Internet-of-Things (IoT), edge computing, and multimedia. However,
selecting appropriate cloud data centers for the applications, from a
large number available locations, is a difficult task. The users need to
consider several different aspects in the data center selection, such as
inter-data center network performance, data transfer costs, and the appli-
cation requirements with respect to the network performance.
This paper proposes a data center clustering mechanism to group
befitting cloud data centers together in order to automate data center
selection task as governed by the application needs. Employing our clus-
tering mechanism, we present four different types of clustering schemes,
with different importance given to available bandwidth, latency, and
cloud costs between pair of data centers. The proposed clustering schemes
are evaluated using a large number of data centers from two major public
clouds, Amazon Web Services, and Google Cloud Platform. The results,
based on a comprehensive empirical evaluation of the quality of obtained
clusters, show that the proposed clustering schemes are very effective in
optimizing data center selection as per the application requirements.
1 Introduction and Motivation
Over the last decade with increasing cloud popularity, centralized cloud data
center deployments have been widely used by a large number of cloud applica-
tions [1]. One major disadvantage of single data center application deployments
is that they are prone to single-point-of-failure in case of data center disrup-
tions, for instance, due to a natural disaster. On the other hand, the emergence
of the Internet-of-Things (IoT), edge computing, and over the Internet multi-
media applications demands services be provided at closest proximity to the
end-users, pushing the need of distributed deployments. To address this prob-
lem, distributed cloud data center application deployments are employed [2],
which enable cloud applications and services to be replicated at geographically-
distributed data center locations [3]. However, it is not a trivial task to select
an appropriate set of distributed cloud data centers for an application. The
https://doi.org/10.1007/978-3-030-15035-8_101
Data Center Clustering for Geographically Distributed Cloud Deployments 1031
distributed data center selection depends on a variety of aspects, including

inter-data center network performance, cloud costs, and the application require-
ments with respect to the data center locations and the network performance.
For instance, it is imperative to select distributed cloud data centers with a
low-latency network connection between them for latency-sensitive applications.
Similarly, cloud data centers with available high bandwidth between them are
appropriate for the applications that rely heavily on data transfers between dis-
tributed application components. In addition, the selection of suboptimal data
centers can also lead to costly data migrations since different cloud providers have
different data transfer rates for transferring to and from different geographical
locations. Thus, it is essential to consider data center characteristics as well as
cloud costs to enable efficient and cost-effective distributed cloud deployments
in the data center selection.
To address distributed data center selection issue, this paper proposes a
cloud data center clustering mechanism to group befitting cloud data centers
together based on the application requirements. The proposed clustering mech-
anism can automate the selection of most appropriate cloud data centers, con-
sidering both the network performance aspects as well as incurred cloud costs.
Using our clustering mechanism, we devised four different data center clustering
schemes, which give different weights to the performance and costs aspects of
the data centers. We evaluated the proposed data center clustering schemes with
15 geographically-distributed data centers from Amazon Web Services (AWS)
and 9 data centers from Google Cloud Platform (GCP). The results, based on
a comprehensive empirical evaluation of the quality of obtained clusters, show
that the proposed clustering schemes are very effective in optimizing data cen-
ter selection as per the application requirements. The key contributions of this
paper include:
• A clustering mechanism to group distributed cloud data centers

• Three different clustering schemes, based on latency, bandwidth, and cost of
data transfer between data centers
• An empirical evaluation of the proposed clustering schemes using
geographically-distributed data centers from two major cloud providers (AWS
and GCP).
The rest of the paper is organized as follows: Sect. 2 introduces the relevant
background, related work, and the employed clustering mechanism. Section 3
presents the experimental design, followed by the evaluation of the experiments
in Sect. 4. Section 5 describes some use cases of our data center clustering. Finally,
Sect. 6 concludes the paper.
2 Cloud Data Center Clustering

Clustering is a process of grouping together a set of objects based on their fea-
tures to maximize the intracluster similarity while minimizing the intercluster
similarity [4]. The objects in the same cluster are more similar to each other
1032 D. Pradhan and F. Zahid
than to the objects from a different cluster as measured by some distance func-
tion (e.g., Euclidean distance). Clustering is widely used in data mining, and the
major applications of clustering include scientific data analysis, document cate-
gorization, pattern recognition, and image processing. Clustering can be used to
obtain interesting patterns and distribution of data directly from data sets with
little or no background knowledge [5]. It is important to note that the quality of
a clustering approach is dependent on the implementation of the approach, e.g.,
the similarity measure used by the approach, selected clustering algorithm and
its required parameters (e.g., number of clusters for partitioning algorithms, such
as k-means), the definition and representation of cluster chosen by the user [6].
There are a number of existing works on data clustering across different
disciplines in the field of data mining [7]. Recently, Devi and Murugaboopathi
[3] proposed a cloud data center clustering algorithm by modeling the cloud data
center as a network graph. Specifically, in [3], distributed cloud data centers are
clustered based on their proximity by considering the degree of the cloud data
centers for cluster center selection and the distance between cloud data centers
by employing Dijkstra’s algorithm. The results were evaluated on a distributed
cloud environment simulated using the C programming language over Ubuntu
14.04 operating system.
As compared to the existing literature, we clustered data centers across dif-
ferent geographical locations from the same cloud providers and across cloud
providers. To compute the similarity between data centers, we used three dif-
ferent metrics: latency, bandwidth, and cost of data transfer. In addition, we
performed experiments on different data centers from real cloud providers for
authentic data.
To cluster the data centers, we employed affinity propagation clustering (AP)
[8] that is based on message passing techniques. Specifically, AP is an exemplar-
based clustering algorithm which finds a set of data points that best exemplify
the data and associates each data point with one exemplar. AP works based on
similarities between pairs of data points (i.e., n * n similarity matrix for n data
points). The detailed description of AP can be checked from [8]. In our context,
we used affinity clustering because AP does not require the number of clusters to
be prespecified unlike other clustering techniques (e.g., k-means [9]) and AP is
a fast clustering algorithm. In addition, we can measure the similarity between
each pair of data centers using the specified metrics (e.g., latency) as required
by AP.
3 Experimental Design
3.1 Research Questions
Our experimental design approach aims at answering the following research ques-
tions.
• RQ1. Do data centers that are close geographically perform better than data
center that are located geographically far?
• RQ2. Can data centers from the same cloud provider in different geograph-
ical locations be clustered together based on the cost of data transfer, and
performance (latency and bandwidth)?
• RQ3. Is it more effective to use data center across different geographical
locations from the same cloud provider or across different cloud providers?
3.2 Selected Data Centers
We selected 24 data centers from two popular cloud providers: Amazon Web
Services (AWS), and Google Cloud Platform (GCP). We used (1) data centers
across all the regions (i.e., 15 regions) in AWS and (2) data centers from GCP in
identical locations (i.e., 9 regions) as AWS. Specifically, we selected data centers
in N. Virginia, Ohio, N. California, Oregon, Mumbai, Seoul, Singapore, Sydney,
Tokyo, Canada (Central), Frankfurt, Ireland, London, Paris, and Sao Paulo for
AWS. In addition, we selected data centers in N. Virginia, Oregon, Mumbai,
Singapore, Sydney, Tokyo, Frankfurt, London, and Sao Paulo for GCP.
3.3 Experiment Settings
We used the tool iperf3 1 to measure the bandwidth across data centers by setting
up data center in one region as a server and the other data centers as a client.
We repeated this process for data centers in all the regions for both AWS and
GCP. In addition, we used ping to measure latency between the data centers.
Experiments were performed using iperf3 and ping for 100 s between each data
centers, and we took an average value for each pair of data center. The time 100 s
was chosen as bandwidth and latency statistics tend to converge around that time
for a given experiment. Moreover, we chose TCP to include the impact of packet
loss in our experiments. To calculate the data transfer cost for datacenter from
the same/different cloud providers, we consulted documentation from AWS2 and
GCP3 . Note that the datacenters have a different cost per data transfer usage
based on the location of data center and if it belongs to the same/different cloud
provider. In our experiments, we took the average cost of data transfer per TB.
Finally, all the values (latency, bandwidth, and cost) were normalized in the
range [0, 1].
The affinity propagation clustering algorithm was implemented from [10].
Regarding the parameter settings for AP, we used the same settings from [10]:
damping factor lambda = 0.5, convergence = 10, maximum number of itera-
tions = 1000. The implementation (using Java and Python) and datasets (i.e.,
latency, bandwidth, and cost between the data centers) are provided online4 .
All the experiments were conducted on the data centers with Ubuntu 16.04 LTS
operating system.
1
https://iperf.fr/.
2
https://aws.amazon.com/ec2/pricing/on-demand/.
3
https://cloud.google.com/compute/pricing#internet egress.
4
https://github.com/Dipsy88/M2EC Data.
Latency (Normalized)
Latency (Normalized)
1.0
1.0
Tokyo Mumbai
0.8
0.8
Singapore
Sydney Sydney
Seoul
0.6
0.6
Tokyo
Sao Paulo
Sao Paulo
0.4
0.4
Singapore
California
Oregon
Mumbai
Toronto
0.2
0.2
Virginia
Virginia
Dublin
0.0
0.0
London
Frankfurt London Frankfurt
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
Bandwidth (Normalized) Bandwidth (Normalized)

(a) AWS (b) GCP
Fig. 1. Data center distances. Note Frankfurt is the base data center.
4 Evaluation
4.1 Data Center Distances
We define data center distance to be a quantitative indicator of the expected

performance between two data centers. Being the two most important criteria for
the network performance, we select latency and bandwidth as our distance met-
rics. Figure 1 presents our data center distance results based on our benchmarks.
The bandwidth and latency values are normalized as described in Sect. 3.3. The
base data center from which these metrics are calculated is Frankfurt. The results
for the AWS cloud are shown in Fig. 1(a). As seen in the figure, the least distant
data center from Frankfurt is London with the best latency and bandwidth. How-
ever, that also results in about 60% drop in the available bandwidth, seen as 0.4
on our normalized scale, compared to the available intra-data center bandwidth
at Frankfurt. As expected, furthest data centers from Frankfurt, such as Sydney
and Seoul, results in worst latency and bandwidth values attained. An impor-
tant finding though is that Singapore though is worse than California when it
comes to the latency but have slightly better bandwidth available to Frankfurt.
On Google Cloud Platform (GCP), as shown in Fig. 1(b), the results are a bit
more surprising. First, the bandwidth available between data centers is quite
low, compared to the AWS. Here, the attained bandwidth between their data
centers in Frankfurt and London is less than 20% of the available intra-data
center bandwidth at Frankfurt. The trend continues to the other data centers
as well. Singapore again is an anomaly considering it shows lower latency than
Sydney (following geographical distance) but slightly a better bandwidth.
Based on our data center distance results, we can answer RQ1 as
geographically-closed data center performs better, as intuition suggests. How-
ever, the data center distance as calculated by latency and bandwidth is not a
direct function of the geographical distance. This is because, different types of
network connections are used between data centers at different cloud providers,
with some being more efficient than others.
Frankfurt, Dublin, London, Mumbai, Paris Frankfurt, Dublin, London, Mumbai, Paris
Toronto, Virginia, Ohio, SaoPaulo Toronto, California, Virginia, Ohio, Oregon, SaoPaulo
250
100 150 200

Latency (ms)
Latency (ms)
California, Oregon, Sydney Seoul, Singapore, Sydney, Tokyo
Seoul, Singapore, Tokyo
150 Base
Base
50
50
0
0
0 50 100 150 0 50 100 150
Bandwidth (Mbits/s) Bandwidth (Mbits/s)

(a) Bandwidth (b) Latency
Frankfurt, Dublin, London, Paris Frankfurt, Dublin, London, Paris

250
Toronto, California, Viginia, Ohio, Oregon Toronto, California, Virginia, Ohio, Oregon
100 150 200

Latency (ms)
Latency (ms)
SaoPaulo, Seoul, Singapore, Sydney, Tokyo Mumbai, Singapore
Seoul, Tokyo
150
Base
Base
50
50
0
0 50 100 150 200 250 0 50 100 150 200 250

(c) Cost (d) All
Fig. 2. Clusters at AWS using different clustering methods
4.2 Clusters
In order to evaluate the quality of the data center clustering we employ, as

detailed in Sect. 2, we measure average bandwidth and latency between data
centers within a cluster, and compare them with the values attained with the
data centers in other clusters. In addition, we employ four different clustering
methods. The first three clustering methods are based on the bandwidth, the
latency, and the cost5 of the data transfer between data centers, whereas the
fourth clustering method employs a weighted function taking all three afore-
mentioned metrics into account with equal weights. The analysis is then done
after selecting a base cluster, which is used to assess the quality of the clusters
obtained. We analyze our clusters’ quality both with a single cloud provider and
multiple cloud providers.
Single Cloud Provider: Figure 2 shows the clusters we obtained at the AWS
using different clustering methods. When clustering is done based on the avail-
able bandwidth between data centers, the impact of the latency is minimal,
as visible in Fig. 2(a). The base cluster constituting data centers from Europe
and India, has on average about 165 Mbps bandwidth available between nodes,
5
Cost-based clustering uses a weighted function equals to 0.75 ∗ Cost + 0.125 ∗
Bandwidth + 0.125 ∗ Latency to avoid ending up with a single big cluster in case of
uniform traffic charges.
which is better than any other inter-cluster bandwidth from the base cluster
(the second best being around 116 Mbps). As latency does not play a role in
clustering, geographically far-away data centers with expected higher latency
between them, such as Oregon and Sydney, are in the same cluster. Similarly,
the latency within one of the clusters formed is actually slightly better than
what is achieved within the base cluster. This indicates that if an application
is geographically-distributed and is not sensitive to latency and require higher
bandwidth, it is quite possible that geographically distant data centers make a
better combination than the nearby data centers. However, when the cluster-
ing employs latency as the considered metric, as shown in Fig. 2(b), clusters are
formed using data centers that are geographically close to each other. Taking
the last example, Sydney is now in a Asia-Pacific cluster, while Oregon lies in a
cluster of American data centers.
Frankfurt, London Frankfurt, London,Virginia, Oregon, SaoPaulo
200
Virginia, Oregon, SaoPaulo Mumbai, Singapore, Sydney, Tokyo
Latency (ms)
Latency (ms)
Mumbai, Singapore, Sydney, Tokyo
150
Base
150
100
100
50
50
Base
0
0 100 200 300 400 500 600 0 50 100 150 200

Frankfurt, London Frankfurt, London

Viginia, Oregon, SaoPaulo Virginia, Oregon, SaoPaulo
100 150 200
Latency (ms)
Latency (ms)
Mumbai, Singapore, Tokyo Mumbai, Singapore, Tokyo

150
100
50
50
Base Base
0
0 100 200 300 400 500 600 0 100 200 300 400 500 600

(c) Cost (d) All
Fig. 3. Clusters at GCP using different clustering methods
Figure 2(c) shows clustering based on the costs. As data transfer pricing at
AWS varies only slightly between the data centers, we see clusters formed based
on a combined weighted function of AWS charges, latency, and bandwidth with
12.5% weight for latency and bandwidth each. For instance, the data transfer
from Europe to locations in the USA is the same ($0.02 per GB), but the data
centers located there are in different clusters considering latency and bandwidth
between the data centers too. Finally, Fig. 2(d) shows clustering based on our
fourth method where each of the cost, latency, and bandwidth are given equal
weights in the weighted function. As shown in the figure, four different clusters
are formed. Compare that with Fig. 2(a) and (b), the data center in Mumbai was
in the cluster with Frankfurt, Dublin, London, and Paris, in both cases. However,
when a weighted method is used which also considers the cost, Mumbai now
constitutes a cluster with Singapore. This type of clustering is very useful for
selecting data centers for a geographically-distributed application that follows a
utility function to assess the importance of cost and performance metrics in the
data center selection.
Our clustering results for the GCP are given in Fig. 3. Thus, we can answer
RQ2 as: it is indeed possible to cluster data centers from different geographical
locations from the same cloud provider.
Frankfurt (AWS,Google), London (AWS,Google) Frankfurt (AWS,Google), London (AWS,Google)

Tokyo (AWS,Google) Virginia (AWS,Google), Oregon (AWS,Google)
250
250
Latency (ms)
Latency (ms)
Mumbai (AWS,Google) SaoPaulo (AWS,Google)
Virginia (AWS,Google) Mumbai (AWS,Google), Singapore (AWS,Google), Tokyo (AWS,Google)
Oregon (AWS,Google) Sydney (AWS,Google)
SaoPaulo (AWS,Google)
Singapore (AWS,Google)
Sydney (AWS,Google)
150
150
50
50
Base Base
0
0 100 200 300 400 0 100 200 300 400

Frankfurt (AWS,Google), London (AWS,Google) Frankfurt (AWS,Google), London (AWS,Google)

Virginia (AWS,Google), Oregon (AWS,Google) Tokyo (AWS,Google)
250
Latency (ms)
Latency (ms)
SaoPaulo (AWS,Google) Mumbai (AWS,Google)

Mumbai (AWS,Google), Singapore (AWS,Google), Tokyo (AWS,Google) Virginia (AWS,Google), Oregon (AWS,Google)
250
Sydney (AWS,Google) SaoPaulo (AWS,Google)

Singapore (AWS,Google)
Sydney (AWS,Google)
150
150
50
50
Base Base
0
0 100 200 300 400 0 100 200 300 400

(c) Cost (d) All
Fig. 4. Clusters when data centers from both AWS and GCP are used
Multiple Cloud Providers: We also ran our clustering experiments based on

the performance benchmarks we obtained using data centers from both AWS and
GCP. As shown in Fig. 4, when both AWS and GCP are used, we get a higher
number of clusters in most cases as compared to when a single cloud provider
is used. The higher granularity represents that the performance gap between
geographically-dispersed data centers is quite high. Taking the clusters based on
the bandwidth (Fig. 4(a)), the average bandwidth within our base cluster (data
centers from Frankfurt and London) is about 400 Mbps, which is more than
four times better than the average bandwidth achieved from the base cluster to
the nodes in the data center of the second best cluster (Virginia). Latency also
follows a similar trend. Moreover, the clusters based on both the bandwidth and
latency, as shown in Fig. 4(a) and (b), respectively, indicate that geographical
location plays a major role in Multi-Cloud data center cluster formation. That
is, the data centers which are in the same city ends up in the same cluster,
irrespective of the cloud provider. Actually, the same trend can be observed on
the clusters formed using all four of our clustering methods, including the cost-
based clusters shown in Fig. 4(c). This is because of the following reasons. First,
when talking about the available bandwidth, even though one can expect that
the bandwidth available between the data centers of the same cloud provider
would be better than the bandwidth available across cloud providers, this is
not particularly true. Thus, we can answer RQ3 as: Available bandwidth across
geographically-dispersed locations is only marginally affected by the choice of the
cloud provider. Similarly, the latency also does not vary considerably based on
the cloud provider, given it is largely determined by the geographical location of
the data centers. For the AWS, the cost of data transfer from AWS data centers
to outside costs considerably higher than within AWS data centers. However, for
the GCP, the charges are the same irrespective of whether the data is transferred
to their own data centers or outside GCP. With the multi-cloud setup using
GCP and AWS, GCP uniform pricing dampens the impact of AWS pricing on
the clusters obtained.
5 Use Cases of Data Center Clustering

In this section we describe how our clustering mechanism can be used to enable
efficient Multi-Cloud application deployments as well as appropriate selection of
data centers for geographically-distributed application based on their require-
ments.
5.1 Geographically-Distributed Application Deployments
Through the last few years, geographically-distributed application deployments

have become an important aspect for various types of cloud applications.
The major advantage is the client performance due to the proximity of the
cloud applications, that helps fuel Internet-of-Things [11], edge computing [12],
and various multimedia application [13]. Moreover, fault-tolerance can also
be an important consideration going in favour of choosing geographically-
dispersed data centers for the cloud applications. However, choosing appropriate
geographically-distributed cloud data centers for an application is not a triv-
ial task as it depends on a variety of aspects include inter-data center network
performance and the application requirements. Some applications are latency-
sensitive and require low latency between the application components deployed,
while others rely heavily on the data transfer, such as file-transfer services. On
top of it, cost is always a concern in the cloud computing. Employing the data
center clustering mechanism we presented in this paper, available cloud data
centers can be clustered based on the application requirements, yielding most
cost-efficient distributed application deployment. Furthermore, our experimen-
tation also gave an insight into the available performance and cost characteristics
between data center of two major cloud providers.
5.2 Cross-Cloud Data Placements and Migrations
An important challenge in Cross-Cloud application deployments is data-

awareness, which refers to the deployments that consider locations of the data
sources into account. Modern cloud applications consume and process data, as
well as generate data to be stored at various on-cloud and off-cloud locations. The
lack of data-awareness may lead to poor application performance and costly data
migrations. For instance, when application deployment decisions do not take in
account the characteristics of the locations of the existing data sources, applica-
tion performance degrades and expensive data migrations are needed. The work
presented in this paper can be used to improve both the data placement and
migration decisions, based on the Cross-Cloud data center clustering. Our data
center clustering mechanism not only takes the performance metrics of the data
centers into account, but also considers the associated costs. Furthermore, the
clustering can easily be extended based on the application characteristics, data
sizes, and performance requirements.
6 Conclusion
This paper proposed a clustering mechanism to group closely-related geographi-

cally distributed cloud data centers together using latency, bandwidth, and cost
of data transfer between the data centers as distance metrics. We empirically
evaluated the quality of the clusters obtained through our clustering mechanism
and found them very effective in optimizing data center selection. The proposed
clustering mechanism can help to automate the selection of most appropriate
cloud data centers for distributed application deployments and data-aware appli-
cation placements.
Acknowledgement. This work has received funding from the European Union’s
H2020 programme under grant agreement no. 731664 (MELODIC).
References
1. Liu, Y., Muppala, J.K., Veeraraghavan, M., Lin, D., Hamdi, M.: Data Center
Networks: Topologies, Architectures and Fault-Tolerance Characteristics. Springer,
Heidelberg (2013)
2. Khethavath, P., Thomas, J.P., Chan-tin, E.: Towards an efficient distributed cloud
computing architecture. Peer-to-Peer Netw. Appl. 10(5), 1152–1168 (2017)
3. Devi, R.K., Murugaboopathi, G.: An efficient clustering and load balancing of

distributed cloud data centers using graph theory. Int. J. Commun. Syst. e3896
(2019)
4. Karypis, G., Han, E.H., Kumar, V.: Chameleon: hierarchical clustering using
dynamic modeling. Computer 32(8), 68–75 (1999)
5. Velmurugan, T.: Performance based analysis between k-Means and Fuzzy C-Means
clustering algorithms for connection oriented telecommunication data. Appl. Soft
Comput. 19, 134–146 (2014)
6. Kaufman, L., Rousseeuw, P.J.: Finding Groups in Data: An Introduction to Cluster
Analysis, vol. 344. Wiley, Hoboken (2009)
7. Jain, A.K., Murty, M.N., Flynn, P.J.: Data clustering: a review. ACM Comput.
Surv. (CSUR) 31(3), 264–323 (1999)
8. Frey, B.J., Dueck, D.: Clustering by passing messages between data points. Science
315(5814), 972–976 (2007)
9. Lloyd, S.: Least squares quantization in PCM. IEEE Trans. Inf. Theory 28(2),
129–137 (1982)
10. Schubert, E., Koos, A., Emrich, T., Züfle, A., Schmid, K.A., Zimek, A.: A frame-
work for clustering uncertain data. Proc. VLDB Endow. 8(12), 1976–1979 (2015)
11. Aazam, M., Khan, I., Alsaffar, A.A., Huh, E.N.: Cloud of Things: Integrating Inter-
net of Things and cloud computing and the issues involved. In: 11th International
Bhurban Conference on Applied Sciences and Technology, IBCAST, pp. 414–419.
IEEE (2014)
12. Shi, W., Dustdar, S.: The promise of edge computing. Computer 49(5), 78–81
(2016)
13. Zhu, W., Luo, C., Wang, J., Li, S.: Multimedia cloud computing. IEEE Signal
Process. Mag. 28(3), 59–69 (2011)
Cost Benefits of Multi-cloud Deployment
of Dynamic Computational Intelligence
Applications
Geir Horn1(B) , Pawel Skrzypek2 , Katarzyna Materka2 ,

and Tomasz Przeździȩk3
1
University of Oslo, P.O. Box 1080, Blindern, 0316 Oslo, Norway
Geir.Horn@mn.uio.no
2
7Bulls.com, Al. Szucha 8, 00-582 Warsaw, Poland
{pskrzypek,kmaterka}@7bulls.com
3
CE-Traffic, a.s., Prvniho pluku 621/8a, 186 00 Praha 8, Czech Republic
tomasz.przezdziek@ce-traffic.com
Abstract. Cost savings is one of the main motivations for deploy-

ing commercial applications in the Cloud. These savings are more pro-
nounced for applications with varying computational needs, like Compu-
tational Intelligence (CI) applications. However, continuously deploying,
adapting, and decommissioning the provided Cloud resources manually
is challenging, and autonomous deployment support is necessary. This
paper discusses the specific challenges of CI applications and provide
calculations to show that dynamic use of Cloud resources will result in
significant cost benefits for CI applications.
1 Introduction
Humans have always been concerned with managing our environments, and
research has made tremendous progress in complexity and cybernetics over the
last century [26]. Approaches based on Model Predictive Control (MPC) have
become the standard way of successfully solving constrained optimal control prob-
lems. However, nature is stochastic, and solving optimal control problems under
uncertainty suffers from the ‘curse of dimensionality’ as the inherent Dynamic
Programming (DP) problem grows exponentially with the dimensionality. In this
case, Stochastic Model Predictive Control (SMPC) [3] has recently appeared as
an alternative, or one must resort to approximation strategies [15] based on Rein-
forcement Learning (RL) [21] or Neuro-Dynamic Programming (NDP) [8], which
uses Neural Networks (NNs) [25] to approximate the objective function.
This work has received funding from the European Union’s Horizon 2020 research
and innovation programme under grant agreement No 731664 MELODIC: Multi-cloud
Execution-ware for Large-scale Optimised Data-Intensive Computing, and from the
European Union’s EUREKA Eurostars research and innovation programme under
agreement No E! 11990 FUNCTIONIZER: Seamless support of serverless applications
in multi-cloud.
https://doi.org/10.1007/978-3-030-15035-8_102
1042 G. Horn et al.
The latter technique is an example of the field of science known as Com-

putational Intelligence (CI) [14]. CI encompasses diverse areas as Fuzzy Logic
(FL) [4], Rough Sets (RS) [31], NN, Evolutionary Computation (EC) [7], Swarm
Intelligence (SI) [22], and the combination of EC and learning [16]; and although
many of these techniques can be used directly to take multi-objective decisions,
e.g. FL [30] and EC [2], the focus here is to use CI to achieve model-free system
awareness, prediction, and control.
CI requires data about the system behaviour in order to be used for prediction
and control, and data about the current operational context in order to create
the desired system awareness. These two aspects are often coupled: The desired
system behaviour depends on its operational context. This means that the CI
algorithms cannot be applied once and for all, but they must be repeatedly re-
applied as the system evolves over time. In the best case, it translates to a phase
of training followed by a phase of exploitation of the trained algorithms. In the
worst case, it is necessary to train the system at the same time as it is being
used for prediction and control, and there could be time constraints on the time
available for the training phase.
Special attention is given in this paper to CI applications based on NNs, as
NNs are elements also in many advanced RL algorithms, e.g. in Deep Q Learning
(DQN) or Double Deep Q-Learning (DDQN) algorithms, and the AlphaGo [23]
and Alpha Zero [24] algorithms. Convolutional Neural Network (CNN) [18]
and variants like Residual Neural Networks (RNNs) [10], which introduces skip
connections between the layers, or Inception [28] with wider convolution fil-
ters, or Transformer [29] algorithms, all contain multiple layers that must be
trained through hundreds of thousands or even millions epochs. The training, by
nature, operates on big datasets that may be fetched from other sources, prepro-
cessed, and transformed before use [19]. Consequently, training NN applications
easily requires massive computing power and special processors like Graphics
Processing Units (GPUs), or even dedicated hardware with machine learning
support [20].
We will argue the case of Cloud computing as a solution in Sect. 2. Two
commercial applications are presented: a financial application in Sect. 3.1, and a
car sharing application in Sect. 3.2, and the cost benefits of using a cross-Cloud
deployment for these applications will be discussed in Sect. 4.
2 The Case for Cross-Cloud Autonomic Deployment

As discussed in Sect. 1, CI systems based on some form of NN require training
significant numbers of models [27], and three typical areas are critical: the pre-
processing of data, the training, and the predictions.
An important part of any CI based system is its ability to pre-process the
input data fast and efficiently [13]. Usually the data is gathered from different
sources, joined, transposed, and pre-processed using advanced algorithms, e.g.
like Fast Fourier transformation [6] or Wavelet transformation [12], normaliza-
tion, and outlier removal. One may also have the training data provided by simu-
lations, and many repetitions of the simulations are needed to get representative
Cost Benefits of Multi-cloud Deployment 1043
sets of input data. Furthermore, one may need many attempts using different
approaches to find the best method for pre-processing data. Pre-processing of
data and training require significant computing power, but usually the execution
time is not strictly limited.
However, the main objective of many CI systems is to generate predictions
in real-time or near real-time [1]. Good examples are the applications discussed
in Sect. 3: financial applications used for investment portfolio optimization, and
dynamic pricing model for vehicle-sharing operators; and many others like inter-
net advertisement or e-commerce retail recommender systems. The predictions
must be calculated very fast and reliably as delayed calculations make the results
obsolete or useless.
Thus, massive computing power is needed for CI applications, and the calcu-
lations are usually done in a distributed way, owing to the high cost of High Per-
formance Computing (HPC) solutions. Cloud resources are usually less expensive
than HPC and more flexible to configure and set up than a HPC infrastructure.
The Cloud computing model allows for very flexible creation and sizing of the
application’s computing infrastructure based on available budget and the Cloud
providers’ offers.
On the other hand, a distributed architecture is usually more complex and
difficult to manage and to optimize than monolithic architectures, and proper
use of cross-Cloud computing, and maximizing its efficiency, are the most impor-
tant challenges for distributed CI applications. A proper computing architec-
ture design with enough performance and reliability is crucial for these applica-
tions. Owing to the distributed architecture, many Cloud providers offer different
options and pricing, and so choosing the optimal solution is very difficult. The
cost savings and performance gains, in terms of number of trained models, pre-
processed data, and number of simulations to run, could be really significant.
However, continuously deploying, adapting, and decommissioning the pro-
vided Cloud resources manually will be challenging. The Multi-cloud Execu-
tion ware for Large scale Optimised Data Intensive Computing (MELODIC)
platform1 offers utility based model driven autonomic cross-Cloud deployment
while respecting the application’s constraints on data placement, timeliness, and
cost [11].
3 Commercial Example Applications
3.1 Financial Trading Solution
Extracting value from available data is becoming the key differentiator in the
financial market. Systems using Artificial Intelligence (AI) can analyse much
more data than humans, and therefore they are expected to react quicker and
make better investment decisions based on more information. The start-up AI
Investments 2 uses AI in a self-improving platform for optimizing an investment
1
http://www.melodic.cloud/.
2
http://www.aiinvestments.pl/.
1044 G. Horn et al.
portfolio in a complete trading solution for the global markets, including a diver-
sified way of signalling transactions, determining market conditions, and man-
aging exposition. Instead of predicting the prices of portfolio assets, the solution
offers “diversified high probability transaction notifications” for exposure, risk,
and position sizing management by applying deep CNN, Long Short-Term Mem-
ory networks [17], as well as an reinforcement learning algorithm based on Monte
Carlo Tree Search with value network [24].
Fig. 1. Data flow and structure of the AI Investments application
Figure 1 presents the application structure:
1. There is a control component for supervising the training: This component is

usually deployed in an active-passive configuration. It orchestrates the work
of the whole system and executes the investment strategies.
2. A communication and data bus based on Kafka3 is used for the whole solution.
3. A set of transformation pipeline components. There are parallel pipelines, one
for each broker, where the data is fetched, transformed, trained, and then the
trained models are stored.
4. Predictions components: These components are used to make a one-time pre-
diction using a given model.
Hence, the application is fully distributed, and both the preprocessing data step
and the training step could be done in a distributed manner. The computing
requirements for this application are:
1. The need for training many Machine Learning (ML) models for pattern recog-
nition, usually with complex architectures [20]. Currently, AI Investments
trains a dedicated model for each market, for each time interval (hourly, daily,
weekly), and for each investment strategy. It means that for 200 markets, 3
time intervals and 5 investment strategies, 3000 models must be trained and
re-trained periodically based on new data.
3
https://kafka.apache.org/.
2. There is a complex preprocessing of each training data set: The data is

currently transposed and joined, and then a Wavelet transformation is
applied [12]. Different Wavelet functions are used depending on the results of
the training, but mostly the Wavelets functions from the Daubechies [5] and
Symlet [5] families are used.
3. Predictions based on the trained networks must be done in near real-time:
One minute is the processing deadline for the shorter time intervals where
data must be gathered from the markets, transposed, transformed, and the
prediction produced.
These requirements imply a high performance architecture for training in terms

of parallel processing and a reliable architecture for prediction in terms of low
latency and availability. On the other hand, the infrastructure costs of the infras-
tructure should obviously be minimized.
The deployment utility aims to minimize costs assuming that the training
of the given number of models should end by a certain deadline. This type
of utility will be used in the production environment for re-training models
used for real-time predictions. The initial time to train one model is estimated
based on the number of models to train, number of epochs needed to train a
model and the average time of training one epoch. Then the number of worker
computers needed is calculated to meet the training deadline. Preferably, on-
premise servers are used and the machines in the Cloud are deployed only if the
on-premise servers are too few to meet the training deadline. The MELODIC
platform [11] will continuously monitor the time it takes to train a model, and
estimate the time remaining to finish the training of all models. Based on that
estimate, MELODIC optimizes the number of worker machines required to meet
the deadline by autonomously adding or removing virtual machines in the Cloud.
For training and experiments, a hybrid infrastructure is used where a part of
the application is deployed in the Cloud and a part of the application is deployed
on-premise. This architecture is forced by very high prices of virtual machines
with GPUs or dedicated hardware for ML training in the Cloud. Also the cost
of using HPC solutions is very high compared with the current hybrid deploy-
ment. The baseline infrastructure for training is therefore on-premise servers
with GPUs, and Cloud resources used only when it is not possible to meet
the time constraints of the training on the local computational resources. Even
assuming that the leading Cloud Providers are equipped with the highest qual-
ity network connections, the communication overhead could be a problem that
must be considered, and work is ongoing to extend MELODIC to take latency
into consideration when assigning the location of data and processing jobs [9].
However, for this application, the time of transferring the data for training the
model is not significant compared with training process: Measurements indicate
that transferring data normally takes up to 5 s, while the training time is over
7 h on average. Hence, the network delay and bandwidth are not bottlenecks for
completing the training.
1046 G. Horn et al.
3.2 Simulation of Car Sharing

Optimali4 is a start-up company which develops the platform for data-driven
optimisation of car sharing services. It applies ML to learn from various data
sources relevant to mobility and deliver optimisation-as-a-service for car sharing
operators helping them to find efficient strategies for fleet relocation. The main
requirements for the deployment of the Optimali platform are:
1. Simulation execution: There is a need to execute many simulations to prepare
data for training of the NNs which are further used to make predictions,
both of the profitability based on the initial position of the vehicles and the
level of satisfied demand. The simulations are Central Processing Unit (CPU)
intensive tasks that could be done in parallel.
2. Monte Carlo tree roll-outs and genetic algorithms Genetic Algorithm (GA)
execution: To find the optimal relocations of the vehicles, a Monte Carlo Tree
Search (MCTS) and GA are used leading to CPU intensive tasks that could
run in parallel. These tasks cover simulation of car sharing rides and solving
vehicle routing problem using vehicle routing solver.
Figure 2 shows the data flow and structure of the Optimali application with the
following elements.
1. The process control component controls the whole process of preparing re-
locations of vehicles including data delivery, simulations, neural networks
training, and Monte Carlo tree search roll-outs.
2. The data delivery component fetches the source data needed for simulations
and calculation of the actual relocations of vehicles.
3. The simulation component executes the simulation of the profitability of the
fleet deployment and satisfied customer demand based on the current place-
ment of vehicles.
Fig. 2. Data flow and the structure of the Optimali application

4
www.optimali.io.
4. The MCTS component performs the MCTS roll-out to find the best set of
relocations. For each roll-out, NNs are used to predict the profitability and
the degree of satisfied customer demand.
5. The GA optimisation component performs the GA based optimization to
find the best set of relocations. For each generation, NNs are used to predict
the profitability and the degree of satisfied customer demand. Alternatively
simulations are performed.
6. Vehicle Routing Problem Solver is used to find optimal routes for the ser-
vice work force to perform neccesary service activities such as refueling or
relocation.
7. Additional worker components for NNs training. These are used for train-
ing each NN model for predicting profitability, and a second NN model for
predicting satisfied demand.
Simulation is the key element of the platform, because the results of the simu-
lation are used for NN training that are further used by the optimization algo-
rithms to find the optimal relocations of the vehicles.
The approach to deployment of the Optimali application is based on similar
deployment utility as for AI Investments application. The goal is to minimize
cost of resources, aiming to finish the process of simulation and prediction in the
given time frame. In this case, the utility based Cloud deployment has significant
impact on the overall work flow. Owing to MELODIC’s dynamic resource alloca-
tion by adding or removing resources, it is possible to meet the time constraints
with the optimal cost almost regardless of the size of the problem instance.
4 Discussion
Reduced cost is the main utility of both applications, and the evaluation is
therefore based on the economic benefits of a cross-Cloud deployment combining
public servers with the alternative scenario of only using private infrastructure
to fulfil the applications’ training requirements.
For the AI Investments application it is assumed that on average 10 models
are trained in parallel, but at peak level up to 50 models could be trained, and
the average number of models can reach 18 models during 120 h per month. The
distribution of the training time is presented on Fig. 3. It presents the percentage
of the models trained by a given time. The figure shows that training time is not
constant and depends of the size of the model, training period length, and the
performance of the server and GPU.
With these parameters two different scenarios were considered.
1. A scenario with 5 on-premise servers each with two GPUs. This corresponds
to two models to be trained per server in the normal average case. Additional
servers with GPUs running in the Cloud must therefore be used as needed.
Based on above assumptions the three years’ Total Cost of Ownership (TCO)
is composed of the following elements:
1048 G. Horn et al.
Fig. 3. Training time distribution for the financial trading solution where a few com-
plicated models require significant time to train and the average training time is 7.1 h.
a. The initial purchase of 5 on-premise servers where each server costs 5 000
USD, giving a total investment of 25 000 USD in total.
b. The cost of running a server and its maintenance covering electricity,
cooling, and repairs is assumed to be 10% of the hardware cost yearly.
Thus the total cost of infrastructure operation is 7 500 USD over the
period of three years’ server depreciation.
c. Cloud resources are used only when needed. Up to 10 models trained
at once could be handled by the private infrastructure, and only if it is
necessary to train more models will the Cloud resources be used. Based on
the parameters described above, the Cloud resources are needed during
the 120 h per month when the average load is higher than normal. On
average 8 additional servers are needed during this time since the Cloud
servers only have one GPU each. The cost of such a Cloud server is around
one USD per hour, so the cost of the used Cloud resources will be 960
USD per month, and 34 560 USD over three years5,6,7 .
Adding the cost elements listed above gives a TCO of 67 060 USD for the
three years of operation. The real cost will be lower as the additional servers
in the Cloud are started on demand by the MELODIC platform and stopped
when no longer needed, hence, the extra servers will be running only when
number of trained models exceed the normal average supported by the private
infrastructure.
2. A scenario with 25 private servers with two GPUs. This corresponds to two
models to be trained per server at peak need. The cost of these servers is
162 500 USD. This is composed of 125 000 USD in hardware cost for the 25
servers, and maintenance cost assumed to be 10% percent of the hardware cost.
5
https://aws.amazon.com/ec2/pricing/on-demand/.
6
https://azure.microsoft.com/pl-pl/pricing/.
7
https://cloud.google.com/compute/pricing.
The difference between the two scenarios is 95 400 USD, so the first scenario
could save almost 60% of the cost of running the application on a private
infrastructure. There are many possible Virtual Machine (VM) types or flavours
provided by the various Cloud Providers. Properly choosing the most optimal
flavour for specific task requires additional calculation. It is done automatically
by MELODIC. The list of the most suitable VM flavours for AI Investments is
depicted in Table 1.
Table 1. Suitable Cloud Virtual machine types from different Cloud providers for the
NN training done by the Financial Trading Solution with prices as available at the
time of writing. Memory is in gigabytes.
VM type VCPUs CPU RAM GPU GPU RAM Price/hour

Amazon Web Services—EU: Ireland—OS: Linux
p2.xlarge 4 61 NVIDIA K80 – 0.972$
g3.4xlarge 16 122 NVIDIA Tesla M60 8 1.210$
Azure—West Europe—OS: Ubuntu Advantage Standard
Nd6 6 122 NVIDIA Tesla P40 24 2.287e
Google Compute Engine—EU: Netherlands
n1-standard-8 8 30 NVIDIA Tesla T4 – 1.3684$
In the table 2 the analysis of the different infrastructure scenarios for AI

Investments is presented. Each row contains one scenario. The private infras-
tructure gives the number of servers each hosting two GPUs providing a certain
GPU time for training the models. According to the above cost assumptions the
cost of this private infrastructure is calculated. If the number of GPUs provided
by the on-premise servers is less than what is needed for the average 10 models
to be trained, the additional GPUs must be sourced from a Cloud base infras-
tructure. Finally, additional GPU resources must be acquired form a Cloud burst
infrastructure when the burst of 50 models must be trained, i.e. 40 models more
than normally handled by the private infrastructure. Finally, the table also shows
the total monthly cost of all servers in the given scenario adding together the
cost of the private infrastructure, the Cloud base and the Cloud burst servers.
Assuming the above maintenance cost, the TCO over a three years’ period is
computed. The cost is minimal for the cases of having 5 private servers, i.e. not
being continuously dependent on using Cloud servers, and only use the Cloud
to manage the bursts.
As a cost saving measure, the models trained in the peak time could be
trained in batches or groups of models instead of training all of them at once.
Each batch is trained at once, and batches are trained sequentially. In this case
the training in peak time is longer, due to sequential training of the models, but
it is more cost effective since less GPUs are used in parallel. However, the batch
computations introduces additional waiting times. Consider a training placed in
1050 G. Horn et al.
Table 2. Different ways to support the parallel training of 10 models in the normal
situation, with occasional bursts of parallel training of 50 models to be trained four
times per month. The cost is in USD.
batch i. It will have to wait for the processing of all earlier batches, which is
on average (i − 1)t̄ where t̄ is the average training of one model. This delay is
suffered by all B models in the same batch, hence the total waiting time for all
models in this batch is (i − 1) · t̄ · B. The total waiting time suffered by all models
under the batched scenario results from adding the waiting times for each of the
n batches together:
n n
1
W = (i − 1) · t̄ · B = t̄ · B (i − 1) = t̄ · B · n(n − 1) (1)
i=1 i=1
2
which follows from the sum of the arithmetic series. Table 3 contains the results
of the cost analysis of the batched processing. It contains only the number of
on premise servers in range [5, . . . , 10], because for the lower number of the on
premises servers would not be used to train models in peak time, they would be
used to train base models.
Similar to the AI Investments, also the Optimali application considers low
cost as the main utility, and the cost of using only a private infrastructure is
compared with running the same number of simulations using virtual machines
in the Cloud at on-demand prices. For the Optimali application it is assumed
that a total of 128 cores are needed to complete the simulation, training, and
optimization work by the deadline. These cores are needed up to two hours
per day, because the optimization of the car sharing fleet is done once per day
and should be done in two hours’ long time slots. With these parameters the
calculations for the two scenarios are as follows.
Table 3. Splitting the models to be trained into batches and executing the jobs in one
batch in parallel, but the batches sequentially may lead to less Cloud servers used, but
also extended time to completion.
Private infrastructure Cloud burst

Number of models in a batch 40 20 10 5 2
Number of batches 1 2 4 8 20
Average execution time (h) 7.1 14.2 28.3 56.6 141.6
Expected additional waits (h) 0.0 141.6 424.8 991.3 2690.6
Servers GPUs TCO over 3 years ($)
5 10 72 141 72 141 72 141 72 141 72 141
6 12 76 659 74 677 70 713 62 785 39 000
7 14 81 177 77 213 69 285 53 428 45 500
8 16 85 695 79 749 67 857 52 000 52 000
9 18 90 213 82 285 66 428 58 500 58 500
10 20 94 731 84 821 65 000 65 000 65 000
1. The scenario with only dynamic Cloud servers: The average cost of these
servers is around 5.5 USD per hour. This is based on using 8 virtual machines
with 16 cores each, where each machine cost 0.68 USD per hour. The monthly
cost of these cloud servers is therefore 330 USD based on 5.5 USD per hour,
two hours per day, and 30 days per month. The TCO over three years is
therefore 11 880 USD, which is 330 USD times 36 months.
2. The scenario using a private infrastructure requires optimally an initial pur-
chase of 11 servers with 12 cores each, as this is the most cost-efficient server
type. The cost of each server is 3 000 USD, and the total cost of the hardware
purchase is 33 000 USD. The additional cost of maintenance is again assumed
to be 10% of the hardware cost, thus amounting to 9 900 USD during the
three years. The TCO is therefore 42 290 USD for the private infrastructure.
The difference between the two scenarios is 31 020 USD, so the Cloud scenario
optimized by MELODIC could save over 70% of the cost of running the appli-
cation on a private infrastructure.
The additional benefits from using automatic cloud deployments with time
constrained utility come from the fact that optimization problems of various size
could be solved in virtually the same time due to the dynamic Cloud scalability
offered by MELODIC and its ability to start as many virtual machines as needed.
The Optimali applications may use many possible VM types or flavours, and
so properly choosing the most optimal flavour requires additional calculations,
similar to the ones shown above for the AI Investments application.
As a next steps of our research, we will investigate other Cloud computing
models. One of the most promising approach is to focus on serverless approach.
1052 G. Horn et al.
We will integrate the results of the Functionizer8 project in terms of deploying,

coordinating, and optimizing serverless components. We will analyse the poten-
tial benefits and the methods to use this approach most effectively. Also, we will
analyse and compare the use of various types of GPU in the Cloud and their
impact on training efficiency.
5 Conclusion
The computational dynamics of CI applications has been investigated and exem-

plified with two real-world commercial applications. The contribution from this
paper is conceptual and intended to show how CI applications can benefit from
cross-Cloud deployments. It is hard to generalise a computational methodol-
ogy suitable for all applications, and it is always possible do everything in the
Cloud and or everything on a private infrastructure. The practical challenge
is, however, to find the right balance with respect to performance and cost, and
MELODIC automates this optimisation process by autonomously allocating and
decommissioning additional computing resources as the demand changes.
We have given examples of the benefits for the two commercial applications
as an inspiration for other similar applications, and shown simple calculations
justifying significant cost benefits for both applications by a hybrid deployment
using both private and public Cloud infrastructures.
References
1. Agrawal, A., Gans, J., Goldfarb, A.: Prediction Machines: The Simple Economics
of Artificial Intelligence. Harvard Business Review Press, Brighton (2018)
2. Zhou, A., Qu, B.Y., Li, H., Zhao, S.Z., Suganthan, P.N., Zhang, Q.: Multiobjective
evolutionary algorithms: a survey of the state of the art. Swarm Evol. Comput.
1(1), 32–49 (2011). https://doi.org/10.1016/j.swevo.2011.03.001
3. Mesbah, A.: Stochastic model predictive control: an overview and perspectives for
future research. IEEE Control Syst. 36(6), 30–44 (2016). https://doi.org/10.1109/
MCS.2016.2602087
4. Bede, B.: Mathematics of Fuzzy Sets and Fuzzy Logic. Studies in Fuzziness and
Soft Computing, vol. 295. Springer, Berlin (2013)
5. Chui, C.K.: Wavelets: A Tutorial in Theory and Applications. First and Second
Volume of Wavelet Analysis and Its Applications (1992)
6. Cochran, W.T., Cooley, J.W., Favin, D.L., Helms, H.D., Kaenel, R.A., Lang,
W.W., Maling, G.C., Nelson, D.E., Rader, C.M., Welch, P.D.: What is the fast
fourier transform? Proc. IEEE 55(10), 1664–1674 (1967). https://doi.org/10.1109/
PROC.1967.5957
7. Fogel, D.B.: What is evolutionary computation? IEEE Spectr. 37(2), 26–32 (2000).
https://doi.org/10.1109/6.819926
8. Bertsekas, D.P., Tsitsiklis, J.N.: Neuro-Dynamic Programming. Optimization and
Neural Computation Series. Athena Scientific, Belmont (1996)
8
https://www.melodic.cloud/functionizer.html/.
Processing (PDP), Pavia, Lombardia/Italy (2019)
10. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition.
CoRR abs/1512.03385 (2015)
11. Horn, G., Skrzypek, P.: MELODIC: utility based cross cloud deployment optimisa-
tion. In: 32nd International Conference on Advanced Information Networking and
Applications (AINA) Workshops, Poland, pp. 360–367. IEEE Computer Society,
Krakow (2018). https://doi.org/10.1109/WAINA.2018.00112
12. Hsieh, T.J., Hsiao, H.F., Yeh, W.C.: Forecasting stock markets using wavelet trans-
forms and recurrent neural networks: an integrated system based on artificial bee
colony algorithm. Appl. Soft Comput. 11(2), 2510–2525 (2011). https://doi.org/
10.1016/j.asoc.2010.09.007. The Impact of Soft Computing for the Progress of Arti-
ficial Intelligence
13. Witten, I.H., et al.: Data Mining: Practical Machine Learning Tools and Techniques
(2016)
14. Kacprzyk, J., Pedrycz, W. (eds.): Springer Handbook of Computational Intelli-
gence. Springer Handbooks. Springer, Berlin (2015)
15. Lee, J.M., Lee, J.H.: Approximate dynamic programming strategies and their appli-
cability for process control: a review and future directions. Int. J. Control Autom.
Syst. 2(3), 263–278 (2004)
16. Zhang, J., Zhan, Z.H., Lin, Y., Chen, N., Gong, Y.J., Zhong, J.H., Chung, H.S.H.,
Li, Y., Shi, Y.H.: Evolutionary computation meets machine learning: a survey.
IEEE Comput. Intell. Mag. 6(4), 68–75 (2011). https://doi.org/10.1109/MCI.2011.
942584
17. Kim, J., El-Khamy, M., Lee, J.: Residual LSTM: design of a deep recurrent archi-
tecture for distant speech recognition. CoRR abs/1701.03360 (2017)
18. LeCun, Y., Bengio, Y.: Convolutional networks for images, speech, and time series.
In: Arbib, M.A. (ed.) The Handbook of Brain Theory and Neural Networks, pp.
255–258. MIT Press, Cambridge (1995)
19. Chen, X.-W., Lin, X.: Big data deep learning: challenges and perspectives. IEEE
Access 2, 514–525 (2014)
20. Abadi, M., et al.: TensorFlow: a system for large-scale machine learning. In: 12th
USENIX Symposium on Operating Systems Design and Implementation (OSDI
2016) (2016)
21. Sutton, R.S., Barto, A.G.: Reinforcement Learning, vol. 9. MIT Press, Boston
(1998)
22. Eberhart, R., Shi, Y., Kennedy, J.: Swarm Intelligence. The Morgan Kaufmann
Series in Artificial Intelligence, 1st edn. Morgan Kaufmann, Burlington (2001)
23. Silver, D., Huang, A., Maddison, C.J., Guez, A., Sifre, L., van den Driessche, G.,
Schrittwieser, J., Antonoglou, I., Panneershelvam, V., Lanctot, M., Dieleman, S.,
Grewe, D., Nham, J., Kalchbrenner, N., Sutskever, I., Lillicrap, T.P., Leach, M.,
Kavukcuoglu, K., Graepel, T., Hassabis, D.: Mastering the game of Go with deep
neural networks and tree search. Nature 529, 484–489 (2016)
24. Silver, D., Schrittwieser, J., Simonyan, K., Antonoglou, I., Huang, A., Guez, A.,
Hubert, T., Baker, L.R., Lai, M., Bolton, A., Chen, Y., Lillicrap, T.P., Hui, F.,
Sifre, L., van den Driessche, G., Graepel, T., Hassabis, D.: Mastering the game of
go without human knowledge. Nature 550, 354–359 (2017)
1054 G. Horn et al.
25. Walczak, S.: Artificial neural networks. In: Mehdi Khosrow-Pour, D.B.A. (ed.)
Advanced Methodologies and Technologies in Artificial Intelligence, Computer
Simulation, and Human-Computer Interaction, vol. 1, pp. 40–53 (2019). Chapter 4
26. Umpleby, S.A., Dent, E.B.: The origins and purposes of several traditions in sys-
tems theory and cybernetics. Cybern. Syst. 30(2), 79–103 (1999). https://doi.org/
10.1080/019697299125299
27. Sun, S.: A survey of multi-view machine learning. Neural Comput. Appl. 23(7),
2031–2038 (2013). https://doi.org/10.1007/s00521-013-1362-6
28. Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S.E., Anguelov, D., Erhan,
D., Vanhoucke, V., Rabinovich, A.: Going deeper with convolutions. CoRR
abs/1409.4842 (2014)
29. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser,
L., Polosukhin, I.: Attention is all you need. CoRR abs/1706.03762 (2017)
30. Pedrycz, W., Ekel, P., Parreiras, R.: Fuzzy Multicriteria Decision-Making: Models,
Methods and Applications. Wiley, Hoboken (2010)
31. Pawlak, Z.: Rough sets. Int. J. Comput. Inf. Sci. 11(5), 341–356 (1982). https://
doi.org/10.1007/BF01001956
An Overview of Multi-cloud Computing
Jiangshui Hong1(B) , Thomas Dreibholz1(B) , Joseph Adam Schenkel2 ,

and Jiaxi Alessia Hu2
1
SimulaMet, Pilestredet 52, 0167 Oslo, Norway
{jiangshui,dreibh}@simula.no
2
Durham University, Stockton Road, Durham DH1 3LE, UK
{j.a.schenkel,jiaxi.a.hu}@durham.ac.uk
Abstract. The purpose of this paper is to provide a brief overview of

cloud computing technologies, particularly with respect to multi-cloud
networks. First, the basics of cloud computing concepts are discussed.
Next we outline some challenges facing cloud computing, and discuss
how multi-cloud systems including multi-clouds, hybrid clouds, federated
clouds, and cross-clouds may be used to deal with some of these issues.
Finally, multi-cloud systems may also be used in conjunction with new
developing technologies such as Big Data and Machine Learning, leading
to exciting innovations. These are reviewed in brief. Our goal is to provide
a modern look at the state of the art in multi-cloud computing and
review open issues in the field. The goal is that this paper will help
the reader to understand challenges facing cloud computing, how multi-
cloud computing addresses some of these issues, and inspire community
excitement at the future integration of multi-cloud platforms with other
novel technologies.
1 Cloud Computing Basics

The National Institute of Standards and Technology (NIST) defines cloud com-
putation as a “model for granting users access to a shared pool of configurable
computational resources including networks, servers, storage, applications, and
services that can be rapidly provisioned (adapted for new services) with lit-
tle management or service provider interaction” [25]. Cloud computing is not a
new technology in itself, per se, but rather a new way to incorporate existing
technologies [26]. In other words, cloud computing is simply a way to remotely
access and manage computing resources [34]. This technology has now been inte-
grated into many businesses, as third-party cloud networks allow firms to focus
on core business elements without the need to worry about maintenance or the
computer infrastructure, essentially turning computation into the “fifth utility”
after water, electricity, gas, and telephony [6]. A major advantage of this tech-
nology is that computational resources can be used far more efficiently, resulting
in quicker results due to economies of scale. [6] illustrates this point by com-
paring the costs of operating one server for 1000 h versus 1000 servers for one
hour. Though the monetary costs may be the same, the speed of 1000 servers
https://doi.org/10.1007/978-3-030-15035-8_103
1056 J. Hong et al.
operating in unison to solve a problem greatly reduces the time needed

for the task.
Clouds have five essential qualities that make them useful to consumers [25].
First of all, they can be remotely accessed at any point by a single individual
without requiring any additional human interaction (on-demand self-service).
Additionally, these networks can be used through various interfaces, includ-
ing mobile phones, laptops, and desktop computers (broad network access).
Cloud Service Providers (CSP) also usually rely on multi-tenant models, where
resources are pooled to serve multiple customers (resource pooling). Another
feature of cloud computing is that these networks can be provisioned easily and
released to suit consumer needs (rapid elasticity). Finally, cloud systems fea-
ture metering capabilities for their services that can be used for transparent
feedback for both provider and user (measured service). These characteristics
make cloud computing attractive for small businesses as no upfront investment
is needed. Additionally, they also reduce operation costs, are easily accessible
and decrease the pressure of having to maintain company networks by outsourc-
ing the infrastructure to providers, allowing a company to focus on core business
elements [26].
Currently there are three cloud computing service models that dominate
the industry: Software as a Service (SaaS), Platform as a Service (PaaS), and
Infrastructure as a Service (IaaS). In SaaS, the consumer can operate a software
on the provider’s cloud system, typically over the Internet [25,26]. In PaaS, the
provider allows the consumer to use the cloud network as a platform for their
own developed or acquired programs. Finally, in IaaS, the consumer is provided
with virtual storage and machines.
End Users Resources Managed at Each Layer

Examples:
Business Applications, Web Services, Multimedia

Google apps,
Facebook, YouTube,
Software as a Saleforce.com
Service (SaaS) Application
Software framework (Java/ Python/ .Net) Microsoft Azure,

Storage (DB/ File)
Platform as a Google AppEngine,
Service (PaaS) Amazon Simple DB/S3
Platforms
Computation (VM) Storage (block) Amazon EC2,

GoGrid,
Flexiscale
Infrastructure
Infrastructure
as a Service (IaaS)
CPU, Memory, Disk, Bandwidth
Data Centres
Hardware
Fig. 1. Cloud computing architecture, based on [26]
The Cloud computing architecture, as shown in Fig. 1, consists of four layers,

beginning with the Hardware Layer. The Hardware Layer, as the name suggests,
An Overview of Multi-cloud Computing 1057
contains the physical hardware including servers, switches, power, cooling sys-
tems, etc. Usually, the Hardware Layer takes the form of data centers harbour-
ing thousands of servers. The next level up is the Infrastructure Layer, which
operates by using virtualisation technologies to partition compute and storage
resources. The Platforms Layer is made up of operating systems and application
frameworks, and is mostly used to decrease the strain of deploying applications
directly into virtual machine containers. Finally, the Application Layer consists
of the cloud computing applications, which refers to the applications (SaaS) that
are built on the lower levels (IaaS and PaaS) of the cloud network. For example,
the applications that are available on Google Play or other app stores deliver
applications that run some of their features in clouds, allowing both ease of access
and lowering the computational cost of the app on the device, as operations can
be performed in the cloud.
It is important to discuss deployment models, and overall there are four ways
in which cloud computing can be deployed. In the private model, the cloud is
provisioned for use by a single organisation with multiple users (e.g. business
consumers). The organisation may own and oversee the system, rely on a third
party for service or some combination of the two. In the community cloud model,
multiple organisations may be responsible for cloud provisioning. As with private
cloud computing, one or more of the organisations using the community cloud
may own and operate it, rely on third parties, or again, some combination of
them. A public cloud is one that is accessible by the general public and, unlike the
previous two types of cloud computing, they are only available on the premises
of the cloud provider (e.g. university campus), while the others may be accessible
off premises. Finally, hybrid cloud systems incorporate two or more of the above
deployment models which remain separate entities. However, they are unified by
an underlying technology that allows for data application portability between
the units [25].
2 Issues with Cloud Computing
Now that the basics of cloud computing have been reviewed, it is possible to
discuss some of the issues facing this technology. Though cloud computing is
useful to businesses as it removes the need for planned provisioning and grants
small businesses the ability to scale their computing needs growing in line with
their business requirements, there are some issues and challenges facing cloud
systems that could be the subject of future research [26]. Though still in its
infancy, the widespread use of cloud computing has allowed many researchers
and businesses to begin to weigh in on potential issues facing the technology.
Security: One subject of concern cited by many authors is that of data security
and privacy [6,15,26,27]. In a study by [15], security concerns were flagged as a
key issue in 66 research papers reviewed, with the next highest issue being infras-
tructure at 46, followed by data management at 15. In this study, practitioners
were also interviewed, expressing additional concerns, however still indicated
1058 J. Hong et al.
that security was a practical issue, highlighting that this is a pressing matter for
many researchers and businesses.
Security issues can be subdivided into several subcategories, including: safety
mechanisms, cloud server monitoring, data confidentiality, and avoiding mali-
cious operations [27]. The authors describe problems pertaining to data stor-
age in cloud computing networks, separating it into the components of data
integrity, data confidentiality, data availability, and data privacy. Maintaining
data integrity involves ensuring that data present within a system cannot be
deleted, modified, or fabricated without authorisation. Data confidentiality is
another issue as currently it can be difficult to protect against both insider
threats and external breaches limiting the utility of cloud networks for sensitive
data storage, such as medical records or government files [33]. Another dimen-
sion of security is data availability, or the degree to which data can be recovered.
Finally, data privacy refers to the ability of an individual or group to privately
and selectively share information only amongst themselves.
Despite these concerns, [6] argues that – given the proper preparation – cloud
computing may actually become more secure than other methods. The authors
explain that any obstacles facing cloud computing are also present in traditional
systems and can be solved using technologies that are already used, such as data
encryption, virtual local area networking, firewalls, and more.
Legal Concerns: Legal issues, while related to security issues, are another
distinct concern [15,27]. Data storage rules and laws vary depending on the
location [6]. Although service level agreements between providers and consumers
have been established, there are currently no standards in place. Adopting cloud
computing storage rules to fit local law may be a challenge in the future, and
will require careful planning and coordination when designing these systems.
Data Management: Perhaps an obvious concern to many is what happens to
the data on cloud computing systems in the event of a catastrophic loss of data.
If, for example, a cloud provider company were to go bankrupt, the data may
become irretrievable. This would present a huge problem for large organisations
that are dependent on data storage [15,33].
Interoperability: [23] considers interoperability, or the “ability of diverse sys-
tems and organisations to work together (inter-operate)” to be the second-
greatest challenge to cloud computing systems after data security and trust
issues. This is crucial, because it would allow users to avoid vendor lock-in,
whereby the user becomes dependent on the CSP because they cannot extract
their services to other platforms or clouds. The lack of standardisation means
that it can be difficult to transfer data between cloud systems.
High Latency: Another issue facing cloud computing is the problem of high
latency, or a lag from when the transfer of data begins following an instruc-
tion for its transfer. In the context of cloud computing, this is caused by the
necessity of the various nodes in the cloud to communicate with each other.
Two solutions have emerged as potential solutions to this problem: fog comput-
ing and edge computing. Fog computing may be defined as “a scenario where
a huge number of heterogeneous (wireless and sometimes autonomous) ubiqui-

tous and decentralised devices communicate and potentially cooperate among
them and with the network to perform storage and processing tasks without
the intervention of third parties.” [31]. Another suggested fog definition is that
proposed by [35]. These authors state that fog computing is “a geographically
distributed computing architecture with a resource pool that consists of one
or more ubiquitously connected heterogeneous devices at the edge of network
and not exclusively seamlessly backed by cloud services. This network collab-
oratively provides elastic computation, storage and communication in isolated
environments to many clients within proximity”. To combat latency [17] sug-
gested a fog-based opportunistic spatio-temporal event processing system which
predicts future query region for moving consumers. It is also important to note
that there are several types of fog networks. Edge computing is somewhat simi-
lar and refers to distributed computing paradigms where computation is largely
performed on edge devices (such as smartphones, and laptops) as opposed to on
the central cloud network itself. Computations are shifted to the periphery of
the system.
Vendor Lock-in: Numerous authors have also discussed the problem of ven-
dor lock-in, the situation where the client becomes dependent on the provider,
making shifting provider costs expensive, as it is complicated by legal issues, or
the technical incapability of transfer due to incompatibility issues. This issue
is related to the interoperability problem, due to the presence of several cloud
systems, many of which are incompatible with one another. [21] explains that
a large part of this issue stems from the lack of standardisation with refer-
ence to operating platforms, Application Programming Interfaces (API), Ser-
vice Level Agreements (SLA) and cloud semantics causing interoperability (see
above) issues. This makes it difficult for a company to shift its resources from
one cloud to another, causing a situation where a company may become depen-
dent on a single provider. Although some businesses have already shifted towards
cloud technology, the vendor problem poses a considerable barrier to others due
to a lack of trust in a single CSP [20]. For instance, if a business were to hire a
CSP, and that provider were to go bankrupt, then the business could potentially
lose all the data on that cloud, making it a considerable risk. Solutions to this
problems are varied, but one example from recent literature is the standardisa-
tion of edge computing [28]. In 2017, the EdgeX Foundry project was launched
with the intent of developing a vendor-neutral framework for Internet of Things
(IoT) computing. Meanwhile the Open-Fog consortium has also been launched
as a collaboration between Princeton University and several notable tech com-
panies including Dell and Microsoft in an attempt to standardise fog comput-
ing. Goal of MELODIC (Multi-cloud Execution-ware for Large-scale Optimised
Data-Intensive Computing) [11,12] is to provide a vendor-independent middle-
ware to base cloud applications on.
1060 J. Hong et al.
3 Multi-cloud Solutions
To help deal with some of the issues mentioned above, the next logical progres-
sion in cloud computing is in multi-cloud computing, or cloud systems that utilise
numerous cloud networks and services simultaneously [13]. In short, multi-cloud
systems simply use more than one CSP, though they come in many subcate-
gories that will be described below. This is similar to the definition of cross-cloud
architectures, defined as systems that span across multiple provisioning bound-
aries [14]. Though the exact distinction between multi-clouds and cross-clouds
is lacking (as described in further detail below), it can be understood that in
multi-clouds, the user or business in question utilizes different cloud services for
different applications in their business. For example, they may store data on a
private cloud, share documents on the Google Cloud platform, and perform data
analysis on yet another cloud. On the other hand, cross-cloud architectures are
designed to make the transfer of data and utilisation of apps across the clouds
more streamlined and cohesive [14].
There are various rationales for using multi-clouds or cross-clouds, many of
which address some of the issues mentioned in the previous section. According
to [24], multi-cloud computing directly addresses 10 key issues, some of which
have been mentioned above. Furthermore, it is important to realise that these
issues are both shared and specific to the parties involved, including individual
customers, company level customers, and the CSPs. The 10 key issues addressed
by multi-cloud computing are:
1. Dealing with peaks in service/resource requests using external ones on

demand;
2. Optimising costs or improving quality of services;
3. Reacting to changes of the offers by the providers;
4. Following constraints such as new locations or laws;
5. Ensuring the high availability of resources and services;
6. Avoiding the dependence on only one external provider;
7. Ensuring backups to deal with disasters or scheduled inactivity;
8. Acting as intermediary;
9. Enhance own Cloud service/resources offers, based on agreements with
others;
10. Consuming different services for their particularities not provided elsewhere.
For example, if a company were to provide an online service, and overesti-

mated how much time the users are spending on their site or application, it may
be beneficial to scale down by shifting the infrastructure from lifelong virtual
machines to ephemeral virtual machines, which charge by the minute instead of
by the hour, and therefore reducing costs. In another case, a company manager
may find that different branches or teams within the company rely on differ-
ent private or public cloud infrastructures. So, consolidating the teams would
require substantial overhauls to the underlying logic. In short, changes to the
predetermined plan pertaining to service provisioning, use, or management are
expected challenges to businesses as they grow, and cross-cloud computing is

necessary to deal with the alterations [13,24].
Another advantage of cross-cloud computing is the avoidance of long-term
commitment to a single CSP, tackling issues of interoperability and vendor lock-
in. Because cross-cloud platforms often rely on communicating between their
different cloud components, they may generate new methods of operability, either
through increased standardisation of systems used, or by devising new ways for
clouds to share data with one another on a more universal level. Developing
technologies that would allow the transfer of services to other CSPs would help
with this problem and allow for greater flexibility within the business [13].
Before proceeding, it should be noted that a significant problem in the field
of cloud computing models is a degree of ambiguity present in the literature. As
mentioned above, standardisation is an issue. However, it also pertains to the
definitions surrounding multi-cloud computing. Some authors, such as [13] cat-
egorise multi-clouds, hybrid clouds, and federation clouds all under cross-cloud
computing. On the other hand, [24] places all of the previous models under “mul-
tiple cloud” computation, differentiating between multi-clouds and federation
cloud frameworks. Most recently, [32] classified hybrid and federation infrastruc-
tures under multi-cloud computing, in line with the classification by [24]. For the
purposes of this document, hybrid clouds, multi-clouds, and federation clouds
will be treated as distinct methods to avoid confusion. Future work may benefit
from more concrete definitions such as those provided by the NIST for cloud
computing as a whole [25].
Multi-cloud: Multi-cloud can be defined as cloud systems in which applications
are hosted as chunks among a heterogeneous network of different cloud. However,
with multi-cloud the components are all unique cloud systems, not deployment
methods, as is the case with hybrid clouds (see below). As discussed above,
some have used “multi-cloud” as an umbrella term to describe some of the other
multi-cloud systems described below, leading to some ambiguity. So, for the
definition presented here, multi-clouds will specifically refer to a cloud system
where multiple different cloud networks are combined for different roles, with the
aim of reducing the necessary trust requirements among the CSPs [24,30]. In
other words, in multi-cloud networks, unique CSPs are used for a single business
or organisation needs, and they may all have differing levels of application as
well as differing SLAs. Again, an example of this may be a company that uses a
small amount of one cloud to store or send documents, another private cloud for
sensitive company data, and yet another for data analysis. All the clouds in this
case may have different SLAs, different costs, and different degrees of utilisation
by the company.
The reasons for using the multi-cloud approach are varied. For example a busi-
ness may seek to use a multi-cloud approach for legal considerations, where the
business may operate on one cloud, but requires a different cloud to legally store
data. An- other reason is to avoid using a single CSP and [9] proposes an architec-
tural framework and principles for Programmable Network Clouds hosting Soft-
ware Defined Networking (SDN) and Network Function Virtualisation (NFV) for
1062 J. Hong et al.
geographically distributed multi-cloud computing environments. Costs and SLA-

aware resource provisioning and scheduling that minimised the operating cost
without violating the negotiated SLAs were investigated with respect to tech-
niques for autonomic and timely NFV composition, deployment and management
across multiple Clouds. The authors strongly suggested for future works to inves-
tigate optimisation techniques that simultaneously optimise the VM/container
placement and traffic consolidation, maximising both SLA satisfaction and
reducing costs [9].
Arguably the most useful and well-researched application of multi-cloud
architecture offer enhanced security. [5,7,18,30] describe several multi-cloud
architectures with a primary focus on security, specifically with medical records.
In this paper, the authors discuss four mechanisms for enhanced security. The
first is application imitation, where the clouds within the multi-cloud essentially
“double-check” each other, as results from one cloud can be compared with
those in a different cloud, ensuring data integrity. Layer-wise application par-
tition allows for a separation between logic and data, protecting against flaws
in logic. The application provider cannot recreate logic entirely from user data
unless it is entirely executed on the cloud users system. Partitioning into seg-
ments refers to splitting the application into smaller segments, which are then
run on different clouds, and this process involves two phases: in the first phase,
trusted private clouds receive small computational parts, while untrusted pub-
lic clouds handle higher loads. In the second phase, the computation is shared
amongst the untrusted clouds. The final mechanism for increasing security is the
distribution of chunks, or simply that data is split amongst the different clouds
within the multi-cloud. In this case, different methods for cryptographic data
segmentation have been trialed for the storage and retrieval of data.
Several models have been suggested for these mechanisms. For example, a
cooperative provable data possession (CPDP) scheme was tested by [36]. How-
ever, it had a problem in that the security parameter π is independent of other
parameters, allowing authentication bypass via forgery of π in the response
sequence [18]. More recently, [18] have expanded on integrity checking meth-
ods using the Co-Check scheme, which is based on Boneh-Lynn-Shacham (BLS)
signature and homomorphic tags [18]. As defined by [8], homomorphism is a
property by which a problem in one algebraic system can be converted to a
problem in another algebraic system. The problem can then be solved and con-
verted back, allowing a third party to view the encrypted information and use
it, if it said party had the right tools. Homomorphic encryption in other words,
is the ability to perform operations or calculations on encrypted data without
the need to decrypt it first. In the Co-Check framework, users do not have to
retrieve the entire data file during the challenge-response and integrity checking
stages of the model (stages II and III). Instead the users generate the chal-
lenges for audition by using parts of the metadata restored at the client side to
prompt the audition efficiency and ensure that malicious CSPs cannot bypass
the check [18]. Other authors have also used homomorphic encryption tools
for multi-cloud security. One of the future research directions for multi-cloud
models will be to experiment with security, as homomorphic encryption is still

in infancy with cloud computing application.
Hybrid Cloud: Hybrid clouds are probably the most common type of multi-
cloud framework, and are defined as a subset of multi-cloud networks in which
private cloud deployment methods are combined with one or more public ones.
Unlike a multi-cloud, the deployment methods are not entirely unique clouds.
Hybrid clouds are useful to businesses aiming to share some of their data across
various user types, yet still desire to keep certain elements confidential [3]. For
example, one area where hybrid clouds are currently being explored is in the
field of medicine, where big data is becoming increasingly available. Recently,
[19] explored security solutions for healthcare data using linear network coding
and re-encryption based on ElGamal cryptography in the form of hybrid app-
roach. To provide security and fault tolerance for cloud storage, the authors con-
sidered a linear network coding mechanism and used the ElGamal re-encryption
scheme to encode the key matrix securely. One of the key distinctions between
hybrid clouds and multi-clouds is the nature of the work performed on the sys-
tems. Because multi-clouds use distinct clouds, they can often be used to handle
a variety of tasks and services for a company. On the other hand, hybrid clouds
are more unified in their framework and usually used to handle one task, such
storing and distributing medical records, in a secure fashion.
Federated Cloud Computing: The key distinction for federated cloud com-
puting is that the CSPs agree to share resources. In this system, CSPs share the
client load under a single umbrella federation. This grants the client access to a
potential catalogue of services and resources available and increases the interop-
erability as well as the portability of applications. [32] gives the example of the
EU-based EGI Federated Cloud, which unites more than 20 cloud providers and
300 data centers. Federated clouds can address the vendor lock-in problem in
that applications. Furthermore, data can be migrated from one cloud to another
more easily than in a pure multi-cloud system, due to the underpinning SLAs.
Federated cloud computing likely benefits smaller providers, as they allow to
cover each other’s weaknesses.
Cross-Cloud Computing: According to [14], a cross-cloud application is one
that utilizes more than one cloud API under a single version of the applica-
tion. Cross-cloud architecture may use either dynamically interchangeable APIs
or non-dynamically interchangeable APIs, so long as the APIs in question are
divergent. The literature referring to cross-cloud systems is sparse when com-
pared to the frameworks mentioned above, and this may be due to the lack
of specificity of definition. Even in the [14] paper referenced above, the author
defined hybrid-clouds, multi-clouds, and federated clouds as subclassifications of
cross-cloud computing, not multi-cloud computing, raising questions as to what
the actual distinction between the two is, and what constitutes a “cross-cloud”.
Regardless of this ambiguity, recent work proposes different cross-cloud archi-
tectures for different ends, and in some cases merges the idea of cross-cloud sys-
tems with some of the other multi-cloud models described above. One idea that
1064 J. Hong et al.
appears to pervade cross-cloud is the notion of the “cloud broker” – an entity

that manages the use, performance, and delivery of cloud services, and negoti-
ates relationships between CSPs and consumers [25]. [2] suggested a broker-based
cross-cloud federation manager as a business model for federated cloud networks.
More recently, cross-cloud computing has been highlighted as a potential solution
for the problem of Big Data management in the IoT [29].
4 Research Directions and Considerations for the Future

In the sections above, we have outlined how multi-cloud computing frameworks
can address some of the issues facing cloud computing as a whole, such as data
security and vendor lock-in avoidance. That being said, there are several other
applications for multi-cloud computing networks. Two of the major themes cur-
rently being explored in the literature are cloud computing and Big Data as well
as Machine Learning. The following, we will describe some of the challenges and
current projects in these fields.
Big Data: Though data management has always been an issue with cloud com-
puting, the looming goal of processing vast quantities of data is one of the major
topics facing current research. The most obvious application of big data is again,
medical, where vast quantities of patient information will need to be stored and
processed over large cloud computing networks. This problem is compounded
when medical wearables are introduced, gathering and sending even more infor-
mation into the relevant cloud networks, and decentralising data sources. This
new interconnectivity between household electronics, smart devices, computers,
vehicle monitoring equipment and any other devices that allow for the collec-
tion and exchange of data is referred to as the IoT, and presents a large hurdle
for data analysis, as the data will likely come from a multitude of sources. Due
to the sheer volume of data that either currently exists or is being generated,
traditional analytic methods cannot keep up [29]. This data must be processed
efficiently, analysed (horizontally and vertically), and transferable between dif-
ferent cloud systems. For example, in the case of medical records, a centralised
system may work well within a nationalised healthcare system. However, they
must also be able to account for peripheral services, such as private practices,
which may use alternative cloud systems. It is important that these private firms
possess the ability to feed medical information into the larger, more mainstream
sources.
One solution to the problem of big data storage proposed in 2012 is the idea
of the “rain cloud”, a multi-cloud model where each member cloud completes
an SLA with other member clouds that allow them to work together when data
gets too large for any single cloud to handle [9]. Hybrid clouds are also thought
to be useful for big data management, as the combination between private and
public deployment allows for confidentiality when different parties require dif-
fering (unequal) levels of access to the available data. In many cases businesses
may require portions of their data and infrastructure to remain behind the cor-
porate firewall due to industry standards, legal regulations, or their own desire
for privacy [22]. Other approaches involve divvying up data amongst clouds
for rapid data storage. In Hadoop, an open source implementation of Google’s
method for data storage, large volumes of data are cut into more manageable
chunks across thousands of computers. A parallelised programming API is then
used to distribute the computations to where the data is located and to aggre-
gate the results. This method has tremendous application for bioinformatics and
genomics. However, the technical savvy and lack of bioinformatics tools that run
in parallel has proven to be a barrier up until very recently [4]. They reviewed var-
ious big data cloud computing frameworks including Hadoop, Spark and Flink.
It will be interesting to see what further improvements will be made and how
multi-cloud computing architectures will integrate these approaches. Finally [29]
also mentions the use of Amazon Elastic Map Reduce, IBM BigInsights, and
Microsoft Azure HDInsight as methods for allowing accessible large-scale data
processing frameworks, namely Apache Hadoop, and Apache Spark.
Machine Learning: Although related to big data, machine learning presents its
own sets of challenges and applications for use with cloud computing technolo-
gies. Given the availability of data stored in clouds (e.g. photographs and videos),
there is a wealth of information that can be processed by machine learning algo-
rithms. Techniques such as virtual machine optimisation for healthcare services
are being trialed and studied, and show promise, but must be trialed with other
diseases to demonstrate consistency [1]. [16] has explored the use of hybrid cloud
storage and machine learning for use in the oil and gas industry, where techni-
cal documents contain valuable information from disciplines like geoscience and
engineering and are in general stored in a unstructured format. To improve data
extraction and utilisation, the authors propose a machine-learning-enabled plat-
form, consisting of a carefully selected sequence of algorithms, developed as a
hybrid cloud container that automatically reads and understands the technical
documents with little human supervision. The user can upload raw data to the
platform, which is stored on a private local server. Structured data is generated
as output, which is pushed through to a search engine that is accessible to the
user via the cloud. This allows for a user to quickly identify the most important
parts of technical documents, automate the extraction or relevant data from the
documents, present the data in a meaningful way for further analysis, and finally,
allows the user to share it easily and port it to other platforms.
Future directions for machine learning and multi-cloud computing research
include a shift towards the wide-scale adoption of auto-tuners, especially for the
SaaS layer of the Cloud. [10] also anticipate the advent of new automated tools
for cloud users to benefit from the experience of other users through partially
automated application builders, automated database sharers, query optimisers,
or smart load balancers and service replicators. In other words, the interface
between cloud, application, and user, is expected to become more streamlined
and intuitive based on what is learned from the user experience. As security
has always been a key concern for cloud networks of any kind, novel machine
learning methods may be used for increased security measures. However, the
authors did not specify what these may look like or on what they would be
based, representing a future area of research application.
1066 J. Hong et al.
5 Conclusion
Cloud computing is a blossoming technology with numerous applications in many
industries, including remote computing and storage. Though vendor lock-in and
cyber security are major concerns hybrid clouds, multi-clouds, and federation
clouds may address some of these problems by providing users with alternatives
in the case of scheduled maintenance, breaches, or shut-downs, though each has
their own benefits and disadvantages. Hybrid systems are easily customised to
a given application, however less transferable, and often used for only one task
whereas multi-clouds and federated clouds are suited for businesses that require
multiple tasks or services. Future work should incorporate multi-cloud paradigms
and combine them with other technologies such as machine learning and big data,
as these technologies can either be used to solve some of the current issues with
cloud computing, such as increasing security, or can be used for entirely new
methods of analysis.
References
1. Abdelaziz, A., Elhoseny, M., Salama, A.S., Riad, A.M.: A machine learning model
for improving healthcare services on cloud computing environment. Measurement
119, 117–128 (2018)
2. Abdo, J.B., Demerjian, J., Chaouchi, H., Barbar, K., Pujolle, G.: Broker-based
cross-cloud federation manager. In: 8th International Conference for Internet Tech-
nology and Secured Transactions (ICITST 2013), London, United Kingdom (2013)
3. Aggarwal, R.: Resource provisioning and resource allocation in cloud computing
environment. Int. J. Sci. Res. Comput. Sci. Eng. Inf. Technol. 3(3), 1040–1049
(2018)
4. Alzain, M.A., Soh, B., Pardede, E.: Big data, Hadoop and cloud computing in
genomics. J. Biomed. Inform. 46(5), 774–781 (2013)
5. AlZain, M.A., Soh, B., Pardede, E.: A survey on data security issues in cloud
computing: from single to multi-clouds. J. Softw. 8(5), 1068–1078 (2013)
6. Armbrust, M., Fox, A., Griffith, R., Joseph, A.D., Katz, R.H., Konwinski, A.,
Lee, G., Patterson, D.A., Rabkin, A., Stoica, I., Zaharia, M.: Above the clouds: a
berkeley view of cloud computing. Technical reports UCB/EECS-2009-28, EECS
Department, University of California, Berkeley (2009)
7. Awasthi, P., Mittal, S., Mukherjee, S., Limbasiya, T.: A protected cloud compu-
tation algorithm using homomorphic encryption for preserving data integrity. In:
Recent Findings in Intelligent Computing Techniques, Singapore (2019)
8. Benzekki, K., Fergougui, A.E., Alaoui, A.E.B.E.: A secure cloud computing archi-
tecture using homomorphic encryption. Int. J. Adv. Comput. Sci. Appl. 7(2), 293–
298 (2016)
9. Buyya, R., Son, J.: Software-defined multi-cloud computing: a vision, architectural
elements, and future directions. In: Proceedings of the 18th International Confer-
ence on Computational Science and Applications (ICCSA) (2018)
10. Buyya, R., Srirama, S.N., Casale, G., Calheiros, R., Simmhan, Y., Varghese, B.,
Gelenbe, E., Javadi, B., Vaquero, L.M., Netto, M.A.S., Toosi, A.N., Rodriguez,
M.A., Llorente, I.M., Vimercati, S.D.C.D., Samarati, P., Milojicic, D., Varela,
C., Bahsoon, R., Assuncao, M.D.D., Rana, O., Zhou, W., Jin, H., Gentzsch,
W., Zomaya, A.Y., Shen, H.: A manifesto for future generation cloud computing:
research directions for the next decade. ACM Comput. Surv. 51(5), 105 (2018)
11. Dreibholz, T.: Big data applications on multi-clouds: an introduction to the melodic
project. Keynote Talk at Hainan University, College of Information Science and
Technology (CIST) (2017)
Processing (PDP), Pavia, Lombardia, Italy (2019)
13. Elkhatib, Y.: Mapping cross-cloud systems: challenges and opportunities. In: Pro-
ceedings of the 8th USENIX Conference on Hot Topics in Cloud Computing, Berke-
ley, United State (2016)
14. Elkhatib, Y., Blair, G.S., Surajbali, B.: Experiences of using a hybrid cloud to
construct an environmental virtual observatory. In: Proceedings of the 3rd Inter-
national Workshop on Cloud Data and Platforms, Prague, Czech Republic (2013)
15. Ghanam, Y., Ferreira, J., Maurer, F.: Emerging issues and challenges in cloud
computing - a hybrid approach. J. Softw. Eng. Appl. 5(11A), 923 (2012)
16. Hernandez, N.M., Lucañas, P.J., Graciosa, J.C., Mamador, C., Caezar, L., Pan-
ganiban, I., Yu, C., Maver, K.G., Maver, M.G.: An automated information retrieval
platform for unstructured well data utilizing smart machine learning algorithms
within a hybrid cloud container. In: First EAGE/PESGB Workshop Machine
Learning (2018)
17. Hong, K., Lillethun, D., Ramachandran, U., Ottenwälder, B., Koldehofe, B.:
Opportunistic spatio-temporal event processing for mobile situation awareness. In:
Proceedings of the 7th ACM International Conference on Distributed Event-based
Systems (2013)
18. Mao, J., Tian, W., Zhang, Y., Cui, J., Ma, H., Bian, J., Liu, J., Zhang, J.: Co-
check: collaborative outsourced data auditing in multicloud environment. Secur.
Commun. Netw. 2017, 13 (2017)
19. Modi, K.J., Kapadia, N.: Securing healthcare information over cloud using hybrid
approach. In: Panigrahi, C., Pujari, A., Misra, S., Pati, B., Li, K.C. (eds.) Progress
in Advanced Computing and Intelligent Engineering. Springer, Singapore (2019)
20. Opara-Martins, J.: Taxonomy of cloud lock-in challenges. In: Khatib, M., Salman,
N. (eds.) Mobile Computing. IntechOpen, Rijeka (2018)
21. Opara-Martins, J., Sahandi, R., Tian, F.: Critical analysis of vendor lock-in and its
impact on cloud computing migration: a business perspective. J. Cloud Comput.
5(1), 4 (2016)
22. Ouyang, C., Moura, M.: A vision of hybrid cloud for big data and analytics. Tech-
nical report, IBM Big Data and Analytic Hub (2017)
23. Petcu, D.: Portability and interoperability between clouds: challenges and case
study. In: Towards a Service-Based Internet. Springer, Berlin (2011)
24. Petcu, D.: Multi-cloud: expectations and current approaches. In: Proceedings of
the International Workshop on Multi-cloud Applications and Federated Clouds.
Prague, Czech Republic (2013)
25. Peter, M., Tim, G.: The NIST definition of cloud computing. Technical report,
National Institute of Standards and Technology Gaithersburg, MD 20899-8930
and U.S. Department of Commerce (2011)
26. Qi, Z.: Cloud computing: state-of-the-art and research challenges. J. Internet Serv.
Appl. 1(1), 7–18 (2010)
27. Sun, Y., Zhang, J., Xiong, Y., Zhu, G.: Data security and privacy in cloud com-
puting. Int. J. Distrib. Sens. Netw. 10(7), 190903 (2014)
1068 J. Hong et al.
28. Taherizadeh, S., Jones, A.C., Taylor, I., Zhao, Z., Stankovski, V.: Monitoring self-
adaptive applications within edge computing frameworks: a state-of-the-art review.
J. Syst. Softw. 136, 19–38 (2018)
29. Taherkordi, A., Zahid, F., Verginadis, Y., Horn, G.: Future cloud systems design:
challenges and research directions. IEEE Access 6, 74120–74150 (2018)
30. Thillaiarasu, N., ChenthurPandian, S.: Enforcing security and privacy over multi-
cloud framework using assessment techniques. In: Proceedings of the 10th Interna-
tional Conference on Intelligent Systems and Control (ISCO), India. IEEE, Coim-
batore (2016)
31. Vaquero, L.M., Rodero-Merino, L.: Finding your way in the fog: towards a compre-
hensive definition of fog computing. SIGCOMM Comput. Commun. Rev. 44(5),
27–32 (2014)
32. Varghese, B., Buyya, R.: Next generation cloud computing: new trends and
research direction. Future Gener. Comput. Syst. 79, 849–861 (2018)
33. Vurukonda, N., Rao, B.T.: A study on data storage security issues in cloud com-
puting. Procedia Comput. Sci. 92, 128–135 (2016)
34. Winans, T.B., Brown, J.S.: Cloud Computing: A collection of Working Papers.
Deloitte LLC, New York (2009)
35. Yi, S., Li, C., Li, Q.: A survey of fog computing: concepts, applications and issues.
In: Proceedings of the 2015 Workshop on Mobile Big Data, China. ACM, Hangzhou
(2015)
36. Zhu, Y., Hu, H., Ahn, G.J., Han, Y., Chen, S.: Collaborative integrity verification in
hybrid clouds. In: 7th International Conference on Collaborative Computing: Net-
working, Applications and Worksharing (CollaborateCom), USA. IEEE, Orlando
(2012)
The 1st Workshop on Artificial
Intelligence and Machine Learning
(AIMAL-2019)
Multi-objective Optimal Power Flow
Using Improved Multi-objective
Multi-verse Algorithm
Muhammad Abdullah1 , Nadeem Javaid1(B) , Annas Chand2 ,

Zain Ahmad Khan2 , Muhammad Waqas1 , and Zeeshan Abbas1
1
2
COMSATS University Islamabad, Abbottabad 22010, Pakistan
https://www.njavaid.com
Abstract. This study proposes an improved multi-objective multi-verse

optimization (IMOMVO) algorithm for solving multi-objective optimal
power flow (MOOPF) problem with uncertain renewable energy sources
(RESs). Cross and self-pollination steps of flower pollination algorithm
(FPA) along with crowding distance and non-dominating sorting app-
roach is incorporated with the basic MOMVO algorithm to further
enhance the exploration, exploitation and for well-distributed Pareto-
optimal solution. To confirm the effectiveness of the proposed IMOMVO
algorithm, modified IEEE 30-bus system with security constraints is uti-
lized by considering the total generation cost and active power loss min-
imization. The simulation results obtained with IMOMVO is compared
with MOMVO, NSGA-II, and MOPSO, which reveals the capability of
the proposed IMOMVO in terms of solution optimality and distribution.
1 Introduction
The optimal power flow (OPF) is considered a vital optimization tool for efficient
operation and planning of the power system. The OPF is used to determine the
control variables at a specific time for the power system operation in order
to optimize (maximize or minimize) different objective functions (OFs) while
assuring the technical feasibility of the solution [1].
Limited reserves of fossil fuels and environmental protection issues have
raised the concern to integrate renewable energy resources (RESs) into the grid.
Wind power plants and solar photovoltaic are amongst the most common RESs.
A lot of research work has been done on classical OPF, which consider only
thermal generators. Classical OPF, which is by itself non-linear and non-convex
and large scale optimization problem because of different OFs and constraints,
however, by incorporating uncertain RESs further increase the complexity of the
problem [2].
The OPF was firstly introduced in 1962 by Carpentier. Since the inception of
OPF, many optimization methods have been introduced to the solution of OPF
https://doi.org/10.1007/978-3-030-15035-8_104
1072 M. Abdullah et al.
problems. These approaches can be divided into classical numerical analysis and
modern metaheuristics algorithms. The classical method includes linear pro-
gramming, quadratic programming, interior point, and dynamic programming
etc. These techniques are inadequate for large-scale problems because they are
comprised of complex, lengthy calculation and they are excessively dependent on
initial guess values. In the past few years, different metaheuristic algorithms have
been developed to overwhelm the deficiencies of the classical methods [2,3]. Some
of these algorithms are; whale optimization algorithm (WOA), particle swarm
optimization algorithm (PSO), moth-swarm algorithm (MSA), modified firefly
algorithm (MODFA), non-dominating sorting genetic algorithm (NSGA), differ-
ential evolution (DE), elephant herding optimization (EHO) technique, Teaching
learning-based optimization (TLBO) technique etc. In any case, due to the incon-
sistency of the OPF solution for different objectives, no heuristic algorithm can
be proven optimal for solving all these objectives, Hence, there is always a room
for new techniques and algorithms that can effectively solve multiple objectives
of OPF.
The OPF problem is considered as single-objective if only one objective func-
tion is reckoned to be optimized. However, OPF problem needs to optimize
conflicting objectives in real world, simultaneously. For solving multi-objective
optimization (MOO) problems, two techniques are used in the literature: a pri-
ori and a posteriori [4]. In the first MOO technique, the MOO problem is con-
verted into a single objective (SO) by assigning weight to different objectives and
aggregating them. The importance of each objective is dictated by the weights
assigned to it, which is generally set by an expert. Solving the problem with this
approach does not require any modification in the algorithm. In posteriori MOO
approach, there is no need to define the weights by an expert. In this method,
different solutions are there for the decision maker as compared to the former
one. There is no need to define weights for each objective by an expert and the
Pareto frontier is approximated, even in a single run [5,6].
Increasing the participation of consumers in the electrical network operations
and accommodation of RESs are the most important features of smart grids. The
role of OPF in smart grids is very vital in order to decrease the transmission,
distribution losses, emission of fossil fuels, and the energy generation cost. Sub-
sequently, decreasing the electrical power consumption and prices, concurrently,
enabling consumers and electrical utilities to regulate the demand [7–11].
2 Problem Formulation of MOOPF

The main goal of solving OPF problems is to optimize different objectives by
adjusting control parameters. In MOOPF problems, two are more objectives are
optimized simultaneously subject to the constraints. The MOOPF mathematical
formulation can be described as below [12,13].
Minimizing:
Nobj

fobj = min fj(x́,v́) , (1)
j=1
Multi-objective Optimal Power Flow 1073
subject to:
gk (x́, v́) = 0 k = 1, 2, . . . , G (2)
hl (x́, v́) 0 l = 1, 2, . . . , L, (3)
th
where, fobj are the OFs to be minimized, gk (x́, v́) is the k equality constraint
(EC) which actually represent the load flow equations [14]. The inequality con-
straints (IC) are expressed by h(x́, v́). x́ and v́ are the vectors of state (dependent)
variables and control (independent) variables respectively [15].
The state variables are:
(1) Real power at swing bus (PG1 ).

(2) Reactive power output of generators (QG ).
(3) Voltage at P-Q (Load) buses (VL ).
(4) Loading of transmission lines (SL ).
The vector of static variables can be written as:
x́ = [PG1 , QG1 , . . . , QGN G , VL1 , . . . , VLN B , SL1 , . . . , SLN T L ]
The Control variables are:
(1) Voltages at generation buses (VG ).

(2) Generation buses active power output (PG ).
Besides these, tap settings of transformer, and compensation of shunt VAR (reac-
tive power) compensators can also be considered as control parameters.
Accordingly, vector v́ can be expressed as:
v́ = [PG2 , . . . , PGN G , VG1 , . . . , VGN , . . . , QC1 , . . . , QCN C , T1 , . . . , TN T ].
2.1 Objective Functions
Two OFs are considered in this paper, i.e. minimization of generation cost and
power loss, whose details are provided subsequently.
2.1.1 First Objective: Minimization of Generation Cost

Total generation cost comprises of the operation cost of thermal generators and
RESs along with penalty and reserve cost, which are described below.
Thermal Power Generator Cost Model
For realistic cost model of thermal generators, valve point effect is considered
[16]. In a thermal generating unit steam is regulated by valves to run the turbine
through a separate nozzle group. To achieve maximum efficiency, every nozzle
group operate at its full output [21].For the required output, the valves are
operated in sequence, which results in the non-continuous cost curve. The cost
model of thermal generators considering valve loading effect is given by [17]:
N
TG
CT G (PT G ) = ái + b́i + dí PT Gi

i=1 (4)

+ dí ∗ sin(éi ∗ (PTmin
Gi − PT Gi ) ,
where, NT G is the number of thermal generators. ái , b́i , and ći are the cost
coefficient of ith generator, producing output power PT Gi , dí and éi are the
coefficients of valve point loading effect. The values of thermal units coefficients
are listed in Table 1.
Table 1. Thermal generators coefficient
Cost coefficients á b́ ć d́ é
GT h1 0 2 0.00375 18 0.037
GT h2 0 1.75 0.0175 16 0.038
GT h8 0 3.25 0.00834 12 0.045
Cost Model of Renewable Energy Sources

Generation from RESs does not require any fuel source for its operation. There-
fore, cost function does not exist, except maintenance cost, if the RESs are owned
by independent system operators (ISO). However, in the case when RESs are
operated by private bodies, the ISO pays to them accordingly to the scheduled
power contractually agreed [18].
The direct cost function of the k th wind farm in terms of scheduled power is
given by [19]:
CW (PW ) = gw PW , (5)
where PW is the scheduled power and gw is the direct cost coefficient of the wind
power unit. Similarly, the direct cost for the solar PV unit with scheduled power
PP V , and cost coefficient hpv is given by:
CP V (PP V ) = hpv PP V . (6)
The third renewable energy resource that is considered in this study is the
combination of wind and run of river small hydro generation plant. The scheduled
power is jointly delivered by wind and hydro generators. The hydro generation
considered in this work is of 5 MW, which is proportional to the river’s flow rate
[20]. The direct cost of the combined wind and hydro plant is described as:
CW H (PW H ) = gwh PW H = gw PW H,w + gh PW H,h , (7)

where, PW H represent the scheduled real power from the combined plant, PW H,w
and PW H,h are the electrical energy contribution from the wind and hydro units
respectively. gw is the direct cost coefficient of wind unit and gh is that of hydro
unit.
As wind energy is stochastic in nature, the actual power produced may be less
or more than scheduled power. Therefore, the ISO must have reserve generating
capability to meet the demand. Reserve cost for the wind unit can be calculated
as [21]:
CRW,i (PW sh,i − PW ac,i ) = krw,i (PW sh,i − PW ac,i )
PW sh,i (8)
= krw,i (PW sh,i − pw,i )fw (pw,i )dpw,i .
0
In the case when the output power of the wind generators is greater then the
scheduled power, the ISOs pay penalty cost if the surplus power is not utilized
by them by reducing the power of thermal generators.
The penalty cost of wind plant can be mathematically described as:
CPW,i (PW ac,i − PW sh,i ) = kpw,i (PW ac,i − PW sh,i )
PW r,i
(9)
= kpw,i (pw,i − PW sh,i )fw (pw,i )dpw,i ,
PW sh,i
where, PW sh,i is the scheduled power, PW ac,i is the available power from wind
plant, PW r,i is the rated power, and fw (pw,i )is the probability density function
of wind power. Similarly the reserve and penalty cost of solar and combined
wind and hydro can be calculated. The values of direct, reserve, and penalty
cost coefficients for RESs are provided in Table 2.
The output power of wind power units is directly proportional to wind speed.
Wind speed probability is predominantly represented by Weibull, probability
distribution function (PDF) [17,21,22].
The power given by PV arrays depends on solar irradiance. The probability
distribution of solar irradiance Gs is represented by lognormal PDF [21,23].
Gumbel distribution correctly represent the river flow rate [24,25]. Summary of
the number of turbines, the rated power output from RESs, and the values of
PDF parameters are listed in Table 3.
2.1.2 Second Objective: Minimization of Real Power Loss

The active power loss in the branches of the power system is unavoidable because
of inherent resistance of lines. The real (active) power loss can be mathematically
written as,
nL
nL

f2 (x, u) = Ploss = Gij [Vi2 + Vj2 − 2Vi Vj cos(δij )] (10)
i=1 j=1,j=i
Where nL is the number of branches, Gij is the conductance of the branch which
linked together the ith and j th bus, and δij is the voltage angle between them.
Vi and Vj are the voltages of bus i and bus j. The magnitude of these voltages
are calculated through newton-Raphson load flow equation.
Table 2. RESs direct, reserve, and penalty cost coefficient
Direct cost coefficient

Wind (at bus number: 5, 13) gw1 = 1.6,
gw1 = 1.75
Solar PV (at bus number 11) hs = 1.6
Small hydro (at bus number 13) gh = 1.5
Reserve cost coefficient
Wind (at bus number 5) krw = 3
Solar PV (at bus number 11) krs = 3
Combined wind and hydro (at bus number 13) krwh = 3
Penalty cost coefficient
Wind (at bus number 5) kpw = 1.5
Solar PV (at bus number 11) kps = 1.5
Combined wind and hydro (at bus number 13) kpwh = 1.5
2.2 System Constraints
MOOPF is a non-linear, non-concave optimization problem, which requires to

satisfy different equality constraints (ECs) and inequality constraints (ICs).
These constraints are described as follow.
2.2.1 Equality Constraints

The ECs in MOOPF comprising of real and reactive power load flow equations
[26], which stipulates that the real and reactive power generation must be equal
to the load demand and power losses in the network. The ECs can be mathe-
matically described as:
NB

P G − PD − V l Vl [Gij cos(δij + βij sin(δij ))] = 0 ∀i N B, (11)
j=1
and
NB

QG − QD − Vl Vl [Gij cos(δij + βij sin(δij ))] = 0 ∀i N B. (12)
j=1
2.2.2 Inequality Constraints

ICs are actually the active, reactive power generation limits, security constraints
of lines, and voltage limits of generation and load buses [17,27].
The constraints of generators are:
Table 3. PDF parameters for uncertain RESs.
Wind power generator Solar PV unit

Total wind Cumulative rated Weibull PDF Cumulative Lognormal
turbines power (MW) parameters rated power PDF
(MW) parameters
25 75 k = 2, c = 9 60 μ = 6,
σ = 0.6
Combined wind+small-hydro
Total wind Rated wind Weibull PDF Small hydro Gumbel PDF
turbines power (MW) parameters rated power parameters
(MW)
15 45 k = 2, c = 10 5 λ = 15,
γ = 1.2
PTmin max
G,i PT G,i PT G,i i = 1, . . . , NT G , (13)
min max
PW G,j PW G PW G , (14)
PPmin max
V,k PP V PP V , (15)
min max
PW HG,j PW HG,j PW HG,j , (16)
Qmin max
T G,i QT G,i QT G,i , i = 1, . . . , NT G , (17)
Qmin max
W G QW G QW G , (18)
Qmin max
P V QP V QP V , (19)
Qmin
W HG QW HG Qmax
W HG , (20)
min max
VG,i VG,i VG,i , i = 1, . . . , NG , (21)
where, PT G,i is the active power generation of ith thermal unit, PW G , PP V , and
PW HG is the real power of wind farm, solar PV array, and combined wind and
hydro generators, respectively. Similarly, QT G,i , QW G , QP V , and QW HG are the
reactive power output of the thermal, wind, solar, and combined wind and hydro
generators. VG,i is the voltage of the ith generation bus.
Security constraints includes the voltage limits of P-Q buses and the trans-
mission lines loading capability, which are described as follow:
min max
VLp VLp VLp p = 1, . . . , NLB , (22)
and
max
Slq Slq q = 1, . . . , Nl . (23)
Equation 22 represents the voltage limits for NLB number of load buses and
the line loading constraint is described in Eq. 23 for Nl branches.
Table 4. Best results of two-objective for Case 1 of 30 bus power system at 500
iteration.
Control variables MOMVO MOPSO NSGAII IMOMVO

Best Best Best Best Best Best Best Best
cost power cost power cost power cost power
loss loss loss loss
PT G2 (MW) 23.6459 23.8473 20.0000 57.9310 20.6786 29.3117 24.8195 31.9106
PW G5 (MW) 42.5207 63.1317 46.3870 75.0000 44.7182 101.0292 44.7238 94.7263
PT G8 (MW) 10.0000 10.1696 10.0000 34.6162 10.1694 48.4191 10.1382 39.9864
PP V11 (MW) 41.4415 54.7171 44.9287 60.0000 43.0685 66.2612 40.5211 68.7966
P(W +H)G1 3 (MW) 36.4134 35.9545 32.5331 48.6378 34.8491 36.4088 32.1291 35.9207
VT G1 (p.u) 1.1000 1.1000 1.1000 1.1000 1.1137 1.0808 1.3422 1.2794
VT G2 (p.u) 0.9500 1.0921 1.0913 1.1000 1.0747 1.0813 1.2014 1.1614
VW G5 (p.u) 1.1000 1.0731 1.0989 1.0995 1.1697 1.0735 1.3208 1.2728
VT G8 (p.u) 1.1000 1.1000 1.1000 1.0958 1.1882 1.1506 1.3124 1.2888
VP V11 (p.u) 1.1000 1.0975 1.0822 1.1000 1.1006 1.1565 1.3652 1.3513
V(W +H)G13 (p.u) 1.1000 1.0780 1.0999 1.0982 20.6786 1.0881 1.4777 1.2716
Cost ($/hr) 776.400 818.017 777.777 942.171 776.882 979.451 770.917 954.880
Power loss (MW) 5.526 3.515 5.275 1.760 5.166 1.376 3.572 0.999
3 IMOMVO Algorithm and Its Application to MOOPF

Problem
Many real-world optimization problems entail optimizing different objectives

simultaneously while meeting the constraints. These objectives are often con-
flicting in nature and the optimal value of all these objectives cannot be achieved
at the same time.
3.1 Basic MOMVO Algorithm
MOMVO optimization algorithm is based on the theory of the existence of multi-

verse, recently proposed by Mirjalili [28]. The inspiration of MOMVO is the inter-
action between universes through black, white, and worm wholes. The objects
from one universe move to another via tunnels (black, white holes), and move
within the universe or even from one universe to another through worm hole
[28,29].
3.2 IMOMVO Algorithm
The PF obtained by the MOMVO algorithm at higher times (iteration) is not

well distributed. Therefore, to further increase the exploration and exploitation
of MOMVO, levy flight and self-pollination steps of flower pollination algorithm
[30], is used to create a new set of universes. After finding the inflation rates of
both sets of universes, they are merged together. The solutions are sorted and
selected on the bases of domination level, and crowding distance approach is
used to maintain the diversity of the non-dominated solution. The solutions are
truncated to the size of the archive to reduce computation time. The new solu-
tions are compared to the stored results in the archive and the better solutions
are replaced in the archive.
4 Simulation Results and Discussions
Modified IEEE 30-bus test system is considered to varify the effectiveness of the
proposed IMOMVO for the solution of MOOPF and its performance is compared
with the results of MOPSO, MOMVO, and NSGA-II optimization algorithm.
The codes of these algorithms are written in the MATLAB environment and
MATPOWER (MATLAB package) is used for power flow calculation. All the
simulations are performed on Intel Core i7, 2.00 GHz processor with an 8 GB
RAM personal computer. For the fair comparison of the proposed IMOMVO
with MOPSO, MOMVO, and NSGAII in terms of obtaining the best PF. The
size of the population (Np = 50) and numbers of iterations are taken the same
for all of these algorithms. The summary of the adapted 30-bus is provided in
Table 5.
Two conflicting objectives, total generation cost minimization, and power
loss minimization are solved simultaneously using proposed IMOMVO and its
performance is compared with other optimization algorithms. The PF obtained
from the simulation of these algorithms for solving MOOPF problem of modified
Fig. 1. Pareto optimum front obtained at 20 iterations.


Table 5. Summary of the modified IEEE-30 bus system.
Item Quantity Details

Buses 30 6 generators buses and 24 load
buses
Branches 41
Thermal generators Buses (T G1 , 3 At bus number: 1,2, and 8. Bus
T G2 , T G3 ) number 1 is a slack bus
Wind generator 1 At bus number 5
Solar PV array 1 At bus number 11
Wind generator+small hydro unit 1 At bus number 13
Control variables 11 The real power of generator buses
except for the slack bus and
voltages of all generator buses
Allowed range of load bus voltage 24 0.95 to 1.05 p.u.
Connected load 283.4 MW, 126.2 MVAr
IEEE 30-bus system is shown in Figs. 1, 2, and 3 for 20, 100, and 500 runs. From
the simulation results, it is obvious that the PF obtained from IMOMVO per-
formed better than MOPSO, MOMVO, and NSGA-II, both at lower and higher
iteration (times) in term of solution optimality and distribution. The values of
the control variables for the best cost, best power loss, and the resultant OFs are
provided in Table 4. The PF obtained by MOMVO is better than MOPSO and
NSGA-II at lower iteration, however, at higher iteration, the PF obtained by
MOMVO is not well distributed. The proposed IMOMVO eliminated the limi-
tation of PF distribution of MOMVO algorithm by incorporating the crowding
distance and non-dominating sorting approaches and displayed the best results
both at lower and higher iteration.
4.1 Conclusion
In this study, IMOMVO algorithm has been proposed and applied to solve
MOOPF problem. The proposed algorithm was successfully implemented on
modified IEEE 30-bus power systems. The simulation results reveal the superi-
ority of the IMOMVO over NSGA-II, MOPSO and MOMVO algorithms. The
IMOMVO eliminated the limitation of PF distribution of MOMVO algorithm
and displayed the better results both at lower and higher iteration. Therefore,
the IMOMVO provides better PF. The PF obtained by multi-objective algo-
rithms helps the decision maker to take a better informed-decision, concerning
the compromise between the conflicting objectives.
References
1. Abdi, H., Beigvand, S.D., La Scala, M.: A review of optimal power flow studies
applied to smart grids and microgrids. Renew. Sustain. Energy Rev. 71, 742–766
(2017)
2. Attia, A.-F., El Sehiemy, R.A., Hasanien, H.M.: Optimal power flow solution in
power systems using a novel Sine-Cosine algorithm. Int. J. Electr. Power Energy
Syst. 99, 331–343 (2018)
3. Bai, W., Eke, I., Lee, K.Y.: An improved artificial bee colony optimization algo-
rithm based on orthogonal learning for optimal power flow problem. Control Eng.
Pract. 61, 163–172 (2017)
4. Branke, J., Kaußler, T., Schmeck, H.: Guidance in evolutionary multi-objective
optimization. Adv. Eng. Softw. 32(6), 499–507 (2001)
5. Tan, K.C.: Advances in evolutionary multi-objective optimization. In: Soft Com-
puting Applications. Springer, Heidelberg, pp. 7–8 (2013)
6. Mirjalili, S., et al.: Salp swarm algorithm: a bio-inspired optimizer for engineering
design problems. Adv. Eng. Softw. 114, 163–191 (2017)
7. Javaid, N., et al.: Energy efficient integration of renewable energy sources in the
smart grid for demand side management. IEEE Access 6, 77077–77096 (2018)
8. Javaid, N., et al.: An intelligent load management system with renewable energy
integration for smart homes. IEEE Access 5, 13587–13600 (2017)
9. Khan, M., et al.: Game theoretical demand response management and short-term
load forecasting by knowledge based systems on the basis of priority index. Elec-
tronics 7(12), 431 (2018)
10. Awais, M., et al.: Towards effective and efficient energy management of single home
and a smart community exploiting heuristic optimization algorithms with critical
peak and real-time pricing tariffs in smart grids. Energies 11(11), 3125 (2018)
11. Ahmad, A., et al.: An optimized home energy management system with integrated
renewable energy and storage resources. Energies 10(4), 549 (2017)
12. Nuaekaew, K., et al.: Optimal reactive power dispatch problem using a two-archive
multi-objective grey wolf optimizer. Expert Syst. Appl. 87, 79–89 (2017)
13. Kumar, A.R., Premalatha, L.: Optimal power flow for a deregulated power system
using adaptive real coded biogeography-based optimization. Int. J. Electr. Power
Energy Syst. 73, 393–399 (2015)
14. Chen, G., et al.: Quasi-oppositional cuckoo search algorithm for multi-objective
optimal power flow. IAENG Int. J. Comput. Sci. 45(2), 255–266 (2018)
15. Ghasemi, M., et al.: Multi-objective optimal power flow considering the cost, emis-
sion, voltage deviation and power losses using multi-objective modified imperialist
competitive algorithm. Energy 78, 276–289 (2014)
16. Abdelaziz, A.Y., Ali, E.S., Abd Elazim, S.M.: Flower pollination algorithm to
solve combined economic and emission dispatch problems. Eng. Sci. Technol. Int.
J. 19(2), 980–990 (2016)
17. Biswas, P.P., Suganthan, P.N., Amaratunga, G.A.J.: Optimal power flow solutions
incorporating stochastic wind and solar power. Energy Convers. Manag. 148, 1194–
1207 (2017)
18. Chen, C.-L., Lee, T.-Y., Jan, R.-M.: Optimal wind-thermal coordination dispatch
in isolated power systems with large integration of wind capacity. Energy Convers.
Manag. 47(18–19), 3456–3472 (2006)
19. Wijesinghe, A., Lai, L.L.: Small hydro power plant analysis and development. In:
2011 4th International Conference on Electric Utility Deregulation and Restruc-
turing and Power Technologies (DRPT). IEEE (2011)
20. Tanabe, R., Fukunaga, A.: Success-history based parameter adaptation for differ-
ential evolution. In: 2013 IEEE Congress on Evolutionary Computation (CEC).
IEEE (2013)
21. Reddy, S.S., Bijwe, P.R., Abhyankar, A.R.: Real-time economic dispatch consider-
ing renewable power generation variability and uncertainty over scheduling period.
IEEE Syst. J. 9(4), 1440–1451 (2015)
22. Reddy, S.S.: Optimal scheduling of thermal-wind-solar power system with storage.
Renew. Energy 101, 1357–1368 (2017)
23. Chang, T.P.: Investigation on frequency distribution of global radiation using dif-
ferent probability density functions. Int. J. Appl. Sci. Eng. 8(2), 99–107 (2010)
24. Mujere, N.: Flood frequency analysis using the Gumbel distribution. Int. J. Com-
put. Sci. Eng. 3(7), 2774–2778 (2011)
25. Cabus, P.: River flow prediction through rainfall runoff modelling with a
probability-distributed model (PDM) in Flanders, Belgium. Agric. Water Manag.
95(7), 859–868 (2008)
26. Ghasemi, M., et al.: Solving non-linear, non-smooth and non-convex optimal power
flow problems using chaotic invasive weed optimization algorithms based on chaos.
Energy 73, 340–353 (2014)
27. Mohamed, A.A.A., et al.: Optimal power flow using moth swarm algorithm. Electr.
Power Syst. Res. 142, 190–206 (2017)
28. Mirjalili, S., et al.: Optimization of problems with multiple objectives using the
multi-verse optimization algorithm. Knowl.-Based Syst. 134, 50–71 (2017)
29. Mirjalili, S., Mirjalili, S.M., Hatamlou, A.: Multi-verse optimizer: a nature-inspired
algorithm for global optimization. Neural Comput. Appl. 27(2), 495–513 (2016)
30. Yang, X.-S.: Flower pollination algorithm for global optimization. In: International
Conference on Unconventional Computing and Natural Computation. Springer,
Heidelberg (2012)
Enhanced Robustness Strategy for IoT
in Smart Cities Based on Data
Driven Approach
Talha Naeem Qureshi1 , Nadeem Javaid1(B) , Suhail Ashfaq Butt2 ,

Wahab Khan3 , Sabir Ali Changazi4 , and Muhammad Mudassar Iqbal4
1
2
Division of Science and Technology, University of Education, Lahore, Pakistan
3
Beijing Institute of Technology, Beijing 100000, China
4
Riphah International University, Islamabad 44000, Pakistan
Abstract. Numerous fields have wide range of applications regarding

Internet of Things (IoT) specially smart cities. Due to increasing day
to day growth in number of applications, there is exponential rise in
requirement of IoT devices. IoT network is becoming complex persis-
tently which brings the significant challenge to the robustness of network
topology. IoT is the backbone of smart cities due to inter connecting
wide range of devices and converting a conventional city to smart city.
To improve network topology robustness against targeted, malicious and
intentional attacks have become a critical issue. To deal with the problem
in this article, Enhanced Angle Sum Operation (EASO) ROSE is pre-
sented. Proposed scheme improves the robustness of network topology
without effecting node degree distribution and scale-free property. Topol-
ogy robustness is achieved by degree difference and angle sum operations.
Extensive simulation results verify that our proposed scheme efficiently
generate scale-free typologies for IoT in smart cities and significantly
improve topology robustness against malicious and targeted attacks.
Keywords: Topology robustness · Data driven · Internet of Things ·

Scale-free networks · Malicious attacks
1 Introduction
Revolutionary advances in telecommunication, information technology, electron-
ics and applications resulted into formation of Internet. Initially it was built
with an aim to connect people. However by 2008, it was found that it has
connected even more devices in comparison to people. Exponential growth of
concept primarily been as Internet of Things (IoT). 50 billion new connections
are predicted to be made till 2020 by IoT [1]. IoT [2,3] is a merger of numer-
ous concepts including fifth generation cellular network, hybrid wireless mobile
https://doi.org/10.1007/978-3-030-15035-8_105
Enhanced Robustness Strategy 1085
networks, heterogeneous wireless sensor networks [4–10], ad-hoc networks, Fly-

ing ad-hoc networks and many more. With the increase in number of devices
and applications in IoT, the network chain is becoming more complicated day
by day. Due to huge amount of data processing by IoT node sensors, leads to
serious challenges towards topology robustness. The IoT is having wide variety
of applications in field of smart cities [11,12] and smart grids [13]. Typically
for monitoring, maintenance, data gathering, health care, transportation net-
works, smart grids and smart home utility large number of sensor nodes are
deployed in smart cities [1]. These nodes collect data and communicate with
each other to provide concept of smart cities [13]. U.S electric grid has massive
quantity of small interconnected synchronous grids. Thousands of power plants
are empowering these small grids [1]. All the interconnected grids are monitored
and maintained by sensors, they gather large amount of data and send to cen-
tral database servers. The grid is exploding into a network of things. It may be
considered as biggest case of IoT [1]. Any failure in network of synchronous grids
and power plants may lead to serious life threatening damage.
In recent years topology robustness of big data servers based smart cities net-
work has become a serious issue [14]. Small world and scale-free are two models
from complex network theory. The major utilization of scale-free model is for
homogeneous network topology modeling. Scale-free typologies have numerous
applications including mobile networks, cooperative networks, social networks
and transportation system networks [15]. Node degree follows a power distribu-
tion law in scale-free model [15]. This leads to presence of mostly low degree
nodes in scale-free networks. Random attacks normally target the low degree
nodes. So, scale-free networks are robust against random attacks. Malicious,
targeted and intentional attack focus high degree nodes. Scale-free networks are
fragile to targeted and malicious attacks [16]. Enhancing robustness of scale-free
typologies against malicious attacks is of vital importance to make it applicable
for critical applications like IoT. Following are main contributions of this article.
A. By using data driven approach from big data servers of smart cities, scheme
for enhancing robustness of scale-free IoT networks is proposed.
B. Proposed enhancement in ROSE [15] will enhance the robustness of scale-free
network for IoT without changing node degree distribution. Proposed EASO
ROSE scheme further enhances the topology towards onion-like structure.
Algorithm proven to be more robust against malicious attacks against all
other schemes including ROSE and Simulating Annealing (SA).
Rest of the paper is arranged as follows. Section 2 is related work in which past
efforts in the same domain is presented. Section 3 explains the applied modeling
strategy of IoT network. Topology robustness matrices are discussed in Sect. 4.
Details of our first topology robustness enhancement scheme is presented as
EASO ROSE in Sect. 5. Evaluation results are shown in Sect. 6. Finally conclu-
sion is extracted in Sect. 7 (Fig. 1).
1086 T. N. Qureshi et al.
2 Related Work
Barabasi and Albert proposed a model named as Barabasi-Albert (BA) model
[17]. To generate power-law distribution of node degree in scale-free networks,
BA model follows the following criteria to achieve scale-free characteristics.
• The nodes are added one by one to the network, at each successive step a
new node joins the network such that m ≤ No . m is number of existing nodes
and No total number of nodes.
• When new node joins the existing network, it chooses node to connect based
on probability that a new node will connect to i node depends on its degree ki .
This means that node having more connections are more likely to be chosen
as next hop for new joining node. This is termed as preferential attachment
and concept is termed as M athewEf f ect.
Small number of nodes are having high degree in scale-free networks. Mali-
cious and targeted attacks are intended to destroy high degree nodes. Failing
high degree node results in removal of large number of edges at the same time.
This leads to fragmentation in the network. Issue can be catered by adding
additional edges between nodes. However, it require additional energy resources,
wireless channels and bandwidth. Hence increases cost and additional edges also
destroy scale-free property of the network. Therefore enhancing topology robust-
ness against malicious attacks requires that initial degree distribution of network
remains unchanged after optimization. Multiple modeling strategies to enhance
the robustness of topology have been proposed by many researchers in recent
years.
Schneider et al. [18] proposed a new metric R to measure robustness of
a network. Largest connected cluster is considered after each attack. Metric
R is defined as number of largest connected nodes to the proportion of total
nodes. Criteria is designed to calculate robustness against malicious and tar-

geted attacks. Criteria also consider the scenario in which network suffers a big
damage without completely collapsing. Metric R is more suitable for scale-free
IoT networks as compare to previous metrics.
SimulatingAnnealing algorithm is proposed by Buesseret et al. [19]. Algo-
rithm was based on Monte Carlo method, that includes swapping edges to resem-
ble onion like structure while keeping the degree distribution constant to improve
robustness. Algorithm also considers inferior configurations to enhance the topol-
ogy robustness. Scale-free network can be optimized by SimulatingAnealing,
however algorithm causes many redundant comparisons along with deceasing
computational efficiency.
A complex network model based on dual properties of both small world and
scale-free network is proposed by Liu et al. [20]. Proposed model is designed for
heterogeneous WSNs and not suitable for IoT networks due to having limited
communication range of IoT nodes.
Qui et al. [15] designed a strategy for scale-free WSNs named as ROSE.
Proposed strategy increases the topology robustness of network by two phase
operation and converting a topology into onion like structure. Designed strategy
used the first node as reference for angle sum operation in second phase. However
in reality onion like structure is referenced by highest degree nodes to form inner
core on onion.
Qui et al. [13] proposed a heuristic optimization genetic algorithm based
strategy for IoT network in smart cities. Proposed scheme make possible solution
sets in form of population. Most robust solution set is selected from population
based on Schneider R criteria. Algorithm is based on randomization and cannot
prevent solution from premature convergence.
3 Scale-Free Model for IoT in Smart Cities

Modeling strategy of scale-free topology construction for IoT in smart cities is
discussed in his section. Due to limitation of communication range and energy of
sensor in IoT, scale-free topology have following two constraints. A. Long com-
munication range of sensors are not possible. B. Sensor node degree cannot be
too large. Due to these limitation preferential attachment property of commonly
used method for constructing scale-free topology, BA model cannot be directly
applied [17]. To simulate, information extraction from big data servers in smart
cities is utilized for constructing scale-free typologies. To overcome limitations of
topology, present study considered dense IoT network with as many neighbors
as possible with in communication range. Due to technological advancements in
sensor range and low cost in future, dense IoT networks will be common.
In order to achieve scale-free characteristics it is of great importance that
communication radius of sensor node should be enough to make links with 50$
of the network nodes. If communication radius is not arbitrary large then high
degree nodes will be scattered in the network, forming fragmentation.
To cope with the network fragmentation and scattered high energy nodes,
we deal each fragment part as a cluster. Based on that high energy node each
cluster is transformed into onion like structure. To construct scale-free topologies

for IoT in smart cities environment, BA [17] model is adopted for initial topology
construction. Following are main attributes for constructing scale-free topology
for IoT. A. Edges are constructed between sensor nodes asynchronously if they
are in communication range. B. Local world of newly added node is composed of
all the nodes with communication range. C. Sensor nodes have limited energy.
If any node form arbitrarily large number of edges, it will have to handle too
high traffic. Hence lot of energy consumption, resulting in failure due to quickly
consumed all energy. If that high degree node fail, the entire topology will get
fragment and whole topology will collapse after some cycles. D. If a newly joined
node connects to a high degree node which has reached the maximum degree, it
will be removed [21]. E. Newly joined node prefer to connects to the high degree
node and chooses its neighbor based on roulette method.
The node connects to neighbors with in its communication range to from local
world. All nodes formed connection based on connection probability PConnect .
We define PConnect for neighbor node i as
n

PConnect(i) = Degree(i) / Degree(j) (1)
j=1
Where Degree(i) is degree of ith node and n is the total number of live
nodes in the network. When a new node joins the network, it gets the neighbor
information through the big data servers in the smart city after obtaining its
own location through Global Positioning System (GPS) circuit. All the sensor
broadcast its location after specific interval of time to big data servers. Big
data servers contains the geographic locations of all the sensors of multiple IoT
networks in smart city.
4 IoT Topology Robustness Metrics
Due to failure of nodes owing to complete energy consumption or some attack,

some edges and nodes are destroyed in IoT network. It results in fragmentation
and separation of initial connected network. In general we termed these attacks
as of following two types.
A. Random Attacks: Random selection and destruction of nodes and edges
resulting in network fragmentation.
B. Targeted or Malicious or Intentional Attacks: Selection of the most impor-
tant node on basis of its degree (high degree node). Removal of critical impor-
tant node which results in destruction of large number of edges. Nodes degree
is reevaluated after removing highest degree node. List of sorting degree node is
again updated. Highest degree from list is again targeted and removed until com-
plete topology is isolated. Process causing fragmentation and failure by attacking
very few critically important nodes.
To deal with the targeted or malicious attacks, nodes are prioritized in our
scheme. Degrees of all nodes forming IoT network is extracted from big data
servers of smart cities. Robustness is calculated by destroying critically impor-

tant nodes one by one. High degree nodes are important as compare low degree
nodes. Robustness is calculating by destroying highest degree node from the
network. This step is performed again and again till the complete network is
fragmented and all nodes are isolated. In some scenarios, more than one nodes
are having highest degree, in that case randomly one highest degree node is
selected. Robustness can be explained by critical fraction of attacks or percola-
tion threshold that lead to completely collapse of system. There are some other
robustness measure except percolation concept that uses shortest path [22–24]
and graph spectrum theory [25].
Threshold qc was used traditionally as a threshold value which is fraction
of critical number of attack after which network completely paralyzes [26]. This
measure ignores the condition in which network receives a major critical damage
without being completely collapsing. Targeted and intentional attack may con-
sists of very limited number of hits. Schneider et al. [18] proposed a new robust-
ness metric based of percolation theory. Scheme is based on attack to largest
component during all possible targeted and intentional attacks. Efficiency of the
scheme is based on criteria which is simple as possible and complex as required.
Schneider R robustness mechanism ensures that network remain operative after
reconstructing edges, as possible before collapsing. We have used this enhanced
robustness metric for evaluating our scheme against malicious, targeted and
intentional attacks. Schneider R works by calculating the maximum number of
sub-graphs in the network and evaluate the intensity of damage based on it.
Robustness metric R is defined as following.
N

R = 1/N s(Q) (2)
Q=1
In the above equation N represents the total number of nodes and s(Q) is
the fraction of largest connecting cluster nodes or nodes with largest connected
subgraphs, means highest degree. 1/N is the normalization factor that ensures
that networks of different sizes can be compared. Range of R for any network
can be varied between 1/N to 0.5 numerical value. Where 1/N limit represents
the start topology and 0.5 represents the full mesh topology with fully connected
graph.
5 Enhanced Angle Sum Operation (EASO) ROSE

ROSE [15] is designed to be processed by centralized system in multi hop archi-
tecture. However, EASO ROSE uses database of big data servers of smart cities
for fetch the geographic location of topology along with neighbor list. EASO
ROSE converts the network to as described by Schneider in onion like structure
by tri phase operation. First phase and second phase are detection of indepen-
dent edges followed by degree difference operation as described by ROSE [15].
Then third phase used proposed enhanced angle sum operation.
5.1 Independent Edges

We consider the scale-free IoT topology as a graph with V vertices and M edges.
Then topology graph G = (V, E), where for total N number of nodes, E belongs
to set of M edges. Independent edges are selected on following criteria: A. Lets
consider two edges eij and ekl , all nodes i, j, k and l should be in communication
range of each other. In order to check that all nodes can form connection with
each other. B. Node i and j is not having any edge with k and l. Same for node
k and l is not having any edge with node i and j as shown in Fig. 2.
Fig. 2. Swapping possibilities of independent edges, (a) Original topology, (b) Swap-
ping possibility 1, (c) Swapping possibility 2
5.2 Degree Difference Operation

ROSE [15] first calculates the degree of all four nodes i, j, k, l from independent
edges. Degrees will be calculated as d1 , d2 , d3 and d4 . As described by ROSE,
we considered that all connection methods for swapping between nodes i, j, k, l.
Degree difference for all types of swaps will be recorded as under by ROSE as
per following equations.
SU B0 = |di − dj | + |dk − dl | (3)
SU B1 = |di − dk | + |dj − dl | (4)
SU B2 = |di − dl | + |dj − dk | (5)
R is checked for all the three available edge swapping connection methods.
Further more ROSE uses parameter p as a threshold to limit the degree difference
operation as per following equation.
|di − dk | + |dj − dl | |di − dl | + |dj − dk |
p = max( , ) (6)
|di − dj | + |dk − dl | |di − dj | + |dk − dl |
The value of p is between range (0, 1). In case when a pair of nodes is having
degree difference less than the initial degree difference then the value of R for
different connection methods are compared. The connection method which fulfill
the condition of p and does not reduce R is accepted.
5.3 Enhanced Angle Sum Operation
In onion-like structure nodes are having edges forming the ring formation. In
addition to the ring formation, there exist some edges that are horizontal with
respect to core of onion, ROSE [15] put these horizontal edges into considera-
tion. Through enhanced angle sum operation, topology is converged to onion-
like structure by using these horizontal edges. To begin with the enhanced angle
sum operation EASO ROSE selects centroid node and take centroid as reference.
According to the Schneider et al. [18] and Zhou et al. [27], onion-like structure
consists of core of highest degree nodes surrounded by rings of low degree nodes
as we move towards outer side of the onion. Through simulation and detailed
analysis, it is found that many times first nodes may not belong to core of highest
degree nodes. To cater for the issue, EASO ROSE fetches the degree list of all
nodes in IoT from big data servers of smart cities. Then calculates the highest
degree nodes. High degree node is calculated by following formula after fetching
node degree list.
Highesti = max(Listnode−degree ) (7)
Then highest degree node is selected as centroid is given by:
Centroid = Highesti (8)
An edge eij between node i and node j is selected as shown in Fig. 3.

SurroundingAngle is calculated in after selection of edge. Surrounding Angle
is defined as a angle eij is horizontal to the centroid as shown in Fig. 3(a). Line
between mid point to edge eij and centroid is drawn for calculation of angle
degree. Mid line will form two angles, one with node i and centroid and other
with node j and centroid. The one of the smaller angle is taken as surrounding
angle α and the larger angle is termed as β as shown in Fig. 3(b). We draw a
line between two nodes opposite to angle α, length of the line is calculated as
lena . Same line for angle β is drawn and lenb is calculated. Length of edge eij is
calculated and termed as lenc . All the sounding angles α for all possible swaps
is calculated on basis of Highesti as following.
len2b(Highesti ) + len2c(Highesti ) − len2a(Highesti )

α = arccos (9)
2 ∗ lenb(Highesti ) ∗ lenc(Highesti )
Angle β is obtained by subtracting angle α from 180o as per equation below.
β = 180o − α (10)
By the same process as for edge eij , we calculate the both angles α and β for
the other edge ekl as shown in Fig. 3(c). For all the possible swap combination
of edges between nodes i, j, k, l we calculate the both angles α and β for both
edges as per Fig. 3(d). The combination with largest α is accepted for possible
swap. For assurance Schneider R is again calculated.
Fig. 3. (a) Geometric details of angle sum operation (b) First edge swap possibility,
(c) Second edge swap possibility, (d) Third edge swap possibility
6 Simulations and Results

After extracting nodes deployment locations from smart cities, we deploy and
simulate IoT network using MATLAB. Cities can become smarter and economic
by adopting circular design [28]. In order to simulate performance over circular
smart city we selected circular region over square with diameter equal to 500 m.
Network of 100 IoT sensors is deployed in restricted circular diameter. To validate
the concept of scale-free topology, communication range of each IoT sensor is
restricted to 200 m. Keeping limited energy constraint of IoT network, edge
density is restricted to 30 neighbors for each sensor. Enhanced variant of ROSE
named EASO-ROSE is proposed. Proposed strategies are compared with ROSE
and Simulating Annealing algorithm. Iterations equal to 100 are performed over
all algorithms under same environment. The results are taken over average of
more than 20 times simulations. EASO-ROSE performed better than ROSE
and Simulating Annealing. Best results obtained by EASO-ROSE, ROSE and
Simulating Annealing is 0.126, 1.120 and 0.077 respectively (Fig. 4).
Fig. 4. Nodes distribution in field

EASO-ROSE performed better than ROSE because of modification in Angle

Sum Operation phase of ROSE. Original ROSE considers 1st node as reference or
centroid and calculates surrounding angles with respect to 1st node. However in
original onion like topology, high degree interconnected nodes formed core. Core
is surrounded by rings of low degree interconnected nodes as we move towards
the outer side of the onion. By reference of high degree core nodes, EASO-ROSE
uses highest degree node as reference or centroid (Fig. 5).
Strategy follows highest degree node as reference along with additional cri-
teria. Based on additional criteria similar degree nodes tend to connect each
other. Enabling convergence of topology towards onion like structure and making
0.13
0.12
Schneider R
0.11
EASO-ROSE
ROSE
0.1
Simulating Annealing
0.09
0.08
0.07
0.06
0 10 20 30 40 50 60 70 80 90 100
Number of Iterations
Fig. 5. Topology robustness R
0.25
EASO-ROSE
ROSE
0.2 Simulating Annealing
Schneider R
0.15
0.1
0.05
1 1.5 2 2.5 3 3.5 4 4.5 5
Number of Nodes
Fig. 6. Comparison between EASO ROSE and other algorithms with different sizes of
network topologies in IoT
topology robust against malicious attacks. Enhancements in EASO ROSE proved

to be most efficient in converging topology robust and onion like, resulting in
best results from all other algorithms (Fig. 6).
7 Conclusion
Considering the practical requirements of smart cities for IoT network, we
present our study. Based on data driven concept from big data servers, we extract
node geographic information and neighbor list. Then initial IoT topology is con-
structed. Enhancements in ROSE is suggested and validated by simulations.
Enhancement proved to improve robustness of IoT network topology.
We have designed enhanced criteria of angle sum operation phase of ROSE.
Initial degrees are unchanged for all nodes during the operation. Hence additional
energy cost of adding edges are saved. The enhancement further converts the
topology into onion like structure and increases robustness against malicious
attacks.
Finally the performance of EASO ROSE is tested over scale-free network
typologies of different sizes and edge densities. Performance is compared with
two algorithms including classic ROSE and simulating annealing. Simulation
results proves that our proposed algorithm enhances the topology robustness of
scale-free networks without disturbing the degree distribution.
References
1. Collier, S.E.: The emerging enernet: convergence of the smart grid with the internet
of things. In: 2015 IEEE Rural Electric Power Conference (REPC) (2015)
2. Stankovic, J.A.: Research directions for the Internet of Things. IEEE Internet
Things J. 1(1), 3–9 (2014)
3. Javaid, N., Sher, A., Nasir, H., Guizani, N.: Intelligence in IoT based 5G networks:
opportunities and challenges. IEEE Commun. Mag. 56(10), 94–100 (2018)
4. Naz, A., Javaid, N., Iqbal, M.M., Ali, M., Imran, M., Khan, Z.A.: TBEENISH:
threshold balanced energy efficient network integrated super heterogeneous proto-
col for WSNs. In: 32nd International Conference on Advanced Information Net-
working and Applications (IEEE AINA) (2018)
5. Butt, S.A., Bakar, K.A., Javaid, N., Gharaei, N., Ishmanov, F.H., Afzal, M.K.,
Mahmood, M.K., Mujahid, M.A.: Exploiting layered multi-path routing protocols
to avoid void hole regions for reliable data delivery and efficient energy management
for IoT-enabled underwater WSNs. Sensors 19, 510 (2019)
6. Mateen, A., Awais, M., Javaid, N., Ishmanov, F., Afzal, M.K., Kazmi, S.:
Geographic and opportunistic recovery with depth and power transmission adjust-
ment for energy-efficiency and void hole alleviation in UWSNs. Sensors 19(3), 709
(2019)
7. Hussain, R., Bouk, S.H., Javaid, N., Khan, A., Lee, J.: Realization of VANET-
based clouds services through named data networking. IEEE Commun. Mag. 58,
168–175 (2018)
8. Sher, A., Khan, A., Javaid, N., Ahmed, S.H., Aalsalem, M., Khan, W.Z.: Void
hole avoidance for reliable data delivery in IoT enabled underwater wireless sensor
networks. Sensors 18, 3271 (2018)
9. Qureshi, T.N., Javaid, N.: Enhanced adaptive geographic opportunistic routing
with interference avoidance assisted with mobile sinks for underwater wireless sen-
sor network. In: International Conference on Frontiers of Information Technology
(2018)
10. Qureshi, T.N., Javaid, N., Khan, A.H., Iqbal, A., Akhtar, E., Ishfaq, M.: Balanced
energy efficient network integrated super heterogenous protocol for wireless sensor
networks. In: International Workshop on Body Area Sensor Networks (BASNet)
(2013)
11. Naz, A., Javaid, N., Qureshi, T.N., Imran, M., Ali, M., Khan, Z.A.: EDHBPSO:
enhanced differential harmony binary particle swarm optimization for demand side
management in smart grid. In: 32nd International Conference on Advanced Infor-
mation Networking and Applications (IEEE AINA) (2018)
12. Qureshi, T.N., Javaid, N., Naz, A., Ahmad, W., Imran, M.: A novel meta-heuristic
hybrid enhanced differential harmony wind driven (EDHWDO) optimization tech-
nique for demand side management in smart grid. In: 32nd International Confer-
ence on Advanced Information Networking and Applications (IEEE AINA) (2018)
13. Qiu, T., Liu, J., Si, W., Han, M., Ning, H., Atiquzzaman, M.: A data-driven robust-
ness algorithm for the internet of things in smart cities. IEEE Commun. Mag.
(2017)
14. Wu, J., et al.: Spectral measure of structural robustness in complex networks. IEEE
Trans. Syst. Man Cybern. Part A Syst. Hum. 41(6), 1244–1252 (2011)
15. Qiu, T., Zhao, A., Xia, F., Si, W., Do, W.: ROSE: robustness strategy for scale-free
wireless sensor networks. IEEE/ACM Trans. Networking 55, 18–23 (2017)
16. Li, R.H., Yu, J.X., Huang, X., Cheng, H., Shang, Z.: Measuring robustness of com-
plex networks under MVC attack. In: Proceedings of the 21st ACM International
Conference on Information and Knowledge Management, New York, NY, USA, pp.
1512–1516, October/November 2012
17. Barabási, A.-L., Albert, R.: Emergence of scaling in random networks. Science
286(5439), 509–512 (1999)
18. Schneider, C.M., Moreira, A.A., Andrade Jr., J.S., Havlin, S., Herrmann, H.J.:
Mitigation of malicious attacks on networks. Proc. Nat. Acad. Sci. USA 108(10),
3838–3841 (2011)
19. Buesser, P., Daolio, F., Tomassini, M.: Optimizing the robustness of scale-free
networks with simulated annealing. In: Proceedings of the 10th International Con-
ference on Adaptive and Natural Computing Algorithms (ICANNGA), Ljubljana,
Slovenia, pp. 167–176, April 2011
20. Liu, L., Qi, X., Xue, J., Xie, M.: A topology construct and control model with small-
world and scale-free concepts for heterogeneous sensor networks. Int. J. Distrib.
Sens. Netw. 10(3), 1–8 (2014). Art. no. 374251
21. Xiao, W., Lin, L., Chen, G.: Vertex-degree sequences in complex networks: new
characteristics and applications. Physica A 437(11), 437–441 (2015)
22. Frank, H., Frisch, I.T.: Analysis and design of survivable networks. IEEE Trans.
Commun. Tech. 18, 501–519 (1970)
23. Latora, V., Marchiori, M.: Efficient behavior of small-world networks. Phys. Rev.
Lett. 87, 198701 (2005)
24. Sydney, A., Scoglio, C., Youssef, M., Schumm, P.: Characterizing the robustness
of complex networks. IJITST 2, 291–320 (2010)
25. Fiedler, M.: Algebraic connectivity of graphs. Czech Math. J. 23, 298–305 (1973)
26. Holme, P., Kim, B.J., Yoon, C.N., Han, S.K.: Attack vulnerability of complex
networks. Phys. Rev. E 65, 056109–056123 (2002)
27. Zhou, M., Liu, J.: A memetic algorithm for enhancing the robustness of scale-free
networks against malicious attacks. Phys. A Statist. Mech. Appl. 410, 131–143
(2014)
28. Prendeville, S., Cherim, E., Bocken, N.: Circular cities: mapping six cities in tran-
sition. Environ. Innov. Societal Transitions 26, 171–194 (2018)
Game-Theoretical Energy Management
for Residential User and Micro Grid
for Optimum Sizing of Photo Voltaic
Battery Systems and Energy Prices
Aqdas Naz1 , Nadeem Javaid1(B) , Abdul Basit Majeed Khan2 ,

Muhammad Mudassar Iqbal3 , Muhammad Aqeel ur Rehman Hashmi4 ,
and Raheel Ahmad Abbasi5
1
2
Abasyn University Islamabad Campus, Islamabad 44000, Pakistan
3
Riphah International University, Islamabad 44000, Pakistan
4
University of Engineering and Technology, Taxila 47050, Pakistan
5
Sardar Bahadur Khan Women University, Quetta 87300, Pakistan
Abstract. There is emerging trend in power system, i.e., energy inter-

net that provides energy production, transmission, storage and utiliza-
tion. Which is used to manage and control energy centrally by using
information and communication technologies. In this paper, coordinated
management of renewable and traditional energy is focused. In proposed
work, storage system is embedded with renewable resources in microgrid,
so that after satisfying users energy requirement, surplus energy can be
stored in battery. Energy management is performed with storage capac-
ity includes cost of renewable resources, depreciation cost of battery and
bidirectional energy transmission. User and microgrid are two players
that are involved in non cooperative game theory. In order to maximize
the payoff of user as well as microgrid, the two stage non cooperative
game theoretic method optimizes battery capacity and prices. Which
are charged by micro grid from user and optimize user energy consump-
tion. The distributed algorithm is proposed to explain nash equilibrium
which ensures Pareto optimality in terms of increasing pay off of both
stakeholder. Furthermore, forecasting algorithm back propagation (BP),
Support Vector Machine (SVM) and Stacked Auto Encoder (SAE) are
used for forecasting historical data related to solar power generation.
Predicted data is, thus used by microgrid in defining energy prices and
battery storage capacity.
Keywords: Game theory · Renewable energy resources · Microgrid ·

User · Nash equilibrium

https://doi.org/10.1007/978-3-030-15035-8_106
1098 A. Naz et al.
1 Introduction
In order to achieve quality, traditional grid is transforming into smart grid.
Smart grid is considered as cyber physical system that contain different phys-
ical system such as energy production, energy distribution and energy usage.
Smart grid also contain several advanced techniques such as advanced metering
infrastructure (AMI), energy management system as well as electrical vehicle
[1]. Smart grid is different from traditional grid in terms of functionality and
it works centrally in bidirectional manner. Customer actively involves in distri-
bution of energy by announcing total energy requirement. Large electrical grid
is divided into small scale grid in smaller geographical areas [2,3,19,20]. Small
scale grids are also known as microgrid. Microgrid generates energy with the
help of renewable resources such as wind energy, hydro power and solar power.
That is the reason that microgrid total amount of energy is not defined as it
depends on the weather and climatic conditions. There can be the case when
microgrid does not produce sufficient energy and it purchase energy from utility
in order to full fill requirement of the users [4]. At the same time, if microgrid
is having surplus energy it will sale it to utility. In return, utility will award it
with certain subsidy. In real world energy management system contain different
kind of errors include, implementation, estimation errors. Such kind of errors
make system completely meaningless. Therefore, it is very important to bring
forecasting of energy generation in account along with its expected error. It helps
in performing contingency planning for better energy management [5,21,22]. In
our proposed work, game theoretic energy management system is proposed. It
contains two level players that includes user and micro grid. Back Propagation
(BP), Support Vector Machine (SVM) and Stacked Auto Encoder (SAE) fore-
casting techniques are used to forecast future solar power generation.
2 Contribution
This paper has following contribution: 1. It considers two stage stackelberg game
in which microgrid is leader and user is follower. Besides, MG may purchase
electricity from utility. If load demand of user exceeds its generation capacity as
shown in Figs. 1 and 2. Two scenarios are considered in which energy manage-
ment with and without GT are discussed. In addition energy storage mechanism
is employed to reduce energy cost of MG. Game theory optimize payoff of both
player by the help of predicted energy generated by Photo voltaic (PV) cell.

Here, proposed system model has been discussed with two stage stackelberg
game theory in smart grid. Single microgrid with multiple residential user N =
{1, 2, ....n}. Microgrid is regarded as supplier of power to user to ensure stability
of user [14–16]. Microgrid is equipped with smart meter to make residential users
schedule, consumption of energy. It also contains PV storage system. PV offers
Game-Theoretical Energy Management for Residential User 1099
power for residential user and charging the battery. It also send surplus energy
to the utility. After receiving utility price policy from information network, user
sends demand to micro grid and eventually that demand will be sent to utility.
In proposed system model. User in this scenario have shiftable and unshiftable
loads [17,18]. One day, as a period is taken for scheduling energy consumption
at user end. K denotes each time slots in a period. Here, the utility accepts
electricity demand from user for each time slots and real time price are sent to
user against each time slot. {P k = pk1 , ...pkj , ....pkm }. Microgrid set its prices to
maximize the pay off as per demand of each user in real time.
3.1 User Energy Consumption and Cost Model

Let us assume N number of user is taken in the scenario with a set of N =
{1, 2, ....N }. Complete day is split into K time slots. ln (k) represent total energy
consumption of user includes shiftable and unshiftable load. Daily energy usage
of load of user is represented by equation in [6] representing scheduling of on
peak hours to off peak hours.
3.2 Microgrid Cost Model

For trading purpose, microgrid and user exchange messages between each other.
Energy parameters between user and microgrid are agreed by both stack holders.
Amount of energy required by each user is xkn and the price that is being charged
by microgrid m is pkm . Total demand of electricity from user must be less or equal
than electricity generated by microgrid as shown under:
Xnk ≤ Gkm ∀k ∈ K (1)
However, energy xm demanded by user from microgrid m must satisfy following

constraint:
K
K
N K
en ≤ xkn , and Gkm + Δm ≥ xkm (2)
k=1 k=1 n=1 k=1
where, Δm shows amount of energy that is required by utility in case production

of renewable resources is not enough for usage of N user. Furthermore, price per
unit energy pkm is finalized by microgrid. Here, Pm k
is price that is charged by
microgrid, therefore, energy that is required by user must fulfill constraint Eqs. 1
and 2.
3.2.1 Solar Power Generation

Mostly, hybrid system where generation plant and renewable resources are source
of energy in order to fulfill energy demand of user are difficult to manage. Specif-
ically if solar panels are placed with each user and consequently surplus energy
is sold back to utility. Centralized mechanism needs to be devised to store
energy and distribute among user and utility on requirement and generation
basis [23,24]. Assumption in proposed model is that generation of solar power
1100 A. Naz et al.
is cheaper than the power that is supplied by utility company. Priorities for
providing solar power at any working time in day light to residential users N
by microgrid is: Foremost priority is to be given to fulfill demand of N users
and to charge the batteries, at second priority, surplus energy is to be sold to
utility for generating revenue by trading. For microgrid, lets assume solar power
generation is en (k) ≥ 0 in time slot k. Solar power generation provides energy
en (k) − eln (k) − ebn (k) = 0 means microgrid does not have surplus energy to
be sold to utility. In case en (k) − eln (k) − ebn (k) ≥ 0 means microgrid is having
surplus power that is generated by solar panel is to be sold. Profit of microgrid
that it generates while selling surplus power to utility is:
K

Un = λs en (k) − eln (k) − ebn (k)) (3)
k=1
CnSP = am (L̂ + Δ) + bmg (L̂ + Δ)2 + cm + F |Δ| (4)
3.2.2 Solar Power Storage System

Solar power requires storage system in order to store surplus energy after fulfill-
ing household load. Storage is considered as indispensable need for solar power
generation. Currently variety of batteries are available in markets [33]. Assump-
tion is that microgrid requires battery capacity y mWh and that may varies
within certain limits:
yε[y l , y u ] (5)
where y l represent lower limits of battery capacity whereas y u represent upper
limits of battery. Daily depreciation cost function is represented as follows:
bat
Cm (y) = λbat y (6)
Microgrid
Solar Power
Solar Panel for
Generation Electricity
forcasting using Generation
historical data
D(x,k)
Battery Storage for
Solar Generated
Solar generated
Power
Power
Residential
Users P(K)

bat
where Cm (y) cost of battery depreciation is represented in form of cents/kWh
and it is also correlated with the material and type of the battery. Battery
depreciation cost is linear increasing function according to the total capacity y
of battery. Apart of battery capacity parameter, there are certain parameters
that are required to be taken under consideration, i.e., charging and discharging
efficiency of battery. Let assume, 0 < ηch < 1 and 0 < ηdisch < 1 shows bat-
tery charging and discharging efficiency. s = [s1 , ..., sk , ..., sK ] represents state of
battery for whole day. Here, battery capacity is also defined therefore inequality
constraint regarding state of battery and capacity of battery is as shown:
0 ≤ sk ≤ z (7)
hkch and hkdisch are binary variable that represents pattern of battery charging
and discharging in each time slot. Battery can either be charged or discharged
at the same time that is shown as:
hkch + hkdisch ≤ 1 (8)
Power state of battery at any time slot k is calculated as follows:

1 ˙l
sk+1 = sk + ηch ebk − b (9)
ηdisch k
1
ebh ≤ kkb − b˙lk (10)
ηdisch
ebn (k) represents energy that is required to charge the battery from solar gener-
ated power. Whereas, bln (k) represents total energy to discharge the battery to
fulfill requires to satisfy user requirement. Total energy that is require to charge
and discharge the battery must be under lower and upper limits of battery, thus,
value of eb (k) and bln (k) must satisfy Eqs. 11 and 12
ebk ≤ hkdisch Bch (11)
bln (k) ≤ hkdisch Bdisch (12)

On the basis energy balance
lk = xk + elk + blk (13)
Equation 13 shows that energy consumption of microgrid that is used by user,

charging the battery and surplus energy is transmitted to utility to generate
payoff.
k
Cm = xkm + Cmbat SP
(y) + Cm + Δm (14)
k
Cm shows total cost of microgrid at time slot k.
1102 A. Naz et al.
3.3 Game Formulation and Analysis

In order to study, communication between microgrid and user, Multi leader and
single follower Stackelberg game has been purposed [5]. Primarily, it is multi
player game, where users being leader decides amount of energy to be demanded
k
from microgrid based on price of microgrid Pm . Whereas, microgrid as a fol-
lower decides prices and storage capacity of battery. In this paper, game theory
presented in [6] is extended.
3.4 Equilibrium Analysis

In multi level non cooperative game, the existence of pure equilibrium solution
is not promised always. Thus, it is required to determine the existence of NE
in proposed algorithm. As a matter of fact, variational equality is proven to
be more socially stable as compared to other NE as studied by [7]. Variational
equality is determine for all customers as discussed in [Demand].
Proposition 1: In case of user n N, every day cost function Un is persistently

differentiable in xn for price pkm and electricity consumption by user xn . There-
fore, strategy space of utility function of user U is a non empty convex compact
convex subset of a Euclidean space.
Proof: Owing to persistent characteristics of the daily cost function

Fn (xn , x−n , pk ). It is continuously differentiable in xn . The Hessian of
Fn (xn , x−n , pk ) is calculated positive semi definite. Consequently, cost func-
tion of user n is convex in xn . Proposition 1 depicts that daily cost function
Fn (xn , x−n , pk ) is continuously differentiable. It is also convex in xn . Owing the
fact, energy cost Cn (xn , x−n ) has continuous quadratic form in context of xn .
Proposition 1 is prerequisite of Proposition 2.
Proposition 2: For ∀n N and time slot k K, the NE of the non cooperative
game exist and it is also unique.
Proof: As per proof mentioned in [storage game theory, Th. 6], owing to the fact
that cost function Fn (xn , x−n , pk ) is convex in xn , the NE of the non cooperative
game exist and it is also unique.
As shown in Fig. 1, system model is shown in detail. Where micro grid
announces its prices on the basis of solar power generation forecasting. In this
paper, forecasting is performed using BP, SVM and SAE in order to deter-
mine forecasting error. It helps in efficiently forecasting energy generation. Thus
energy management will be more affective.
4 Simulation and Discussion

In proposed scenario, one year duration solar power data is taken form the NREL
site. In this section, game theoretic energy management effectiveness is verified.
In order to perform analysis on accuracy of solar power prediction. Figure 2
shows solar power generation in single day. It clearly shows the rise and fall
of the generation as per the solar radiations available in different time slots of
the day. Figure 3 shows the trend of user energy consumption which helps in
deciding the peak hours, off peak hours and mid peak hours respectively. N
users are considered in the scenario which contain both shift able and fixed load
and power consumption of users are shown in Fig. 3. Each day is divided into
24 h time slots k. Consumption of electricity usage at each slot varies depending
upon peak hours and off peak hours.
An analytical scenario is presented to verify the performance and effectiveness
of proposed game theory among user and micro grid. Where N user increases
its pay off by adjusting energy consumption. Besides, micro grid optimizes its
payoff by optimally finalizing energy rates pkm and size of the battery y.
In this paper, TOU pricing scheme is adapted and single day is divided in 24
time slots k. Further, entire day is divided into 3 chunks. Single day is divided
into chunks on the basis of energy consumption, i.e., peak hours, off peak hours
and mid peak hours. Each chunk contain different pricing parameters.
DSM is not discussed in this paper on individual basis. Energy consumption
by user N is catered collectively. However, energy management at micro grid
is discussed in detail that includes energy prices pkm and size of the battery
y. Micro grid is equipped with renewable resources, i.e., photo voltaic cells. In
current scenario, it is assumed that micro grid consist of photo voltaic power
cell in terms of renewable resources.
Fig. 2. Solar power generation in 24 h
Fig. 3. Consumption of users in 24 h

1104 A. Naz et al.
Figure 4 shows energy distribution by micro grid in K time slots without

game optimization. Energy is retrieved from utility in hours when solar energy
is not available for generation of solar energy whereas rest of hours, energy is
optimally used using solar power generation and battery. Figure 5 shows energy
distribution by microgrid with game optimization.
Fig. 4. PV energy distribution without game
Fig. 5. PV energy distribution with game
Fig. 6. MAPE of different forecasting model with PV forecasting
Microgrid can take more benefit solar power when there are more day light
hours. It also affect battery capacity optimization once microgrid ensures energy
supply to user. However, day light timings and intensity will not be affect the pro-
posed game theoretic energy management. Microgrid employs 2 MW solar power
generation. Energy output by solar cells per day is shown in Fig. 2. Solar power is
generated during 0600 to 1700 h. User cost relies only on discharging time and it
counts minimal except when battery discharges at peak demand hours. It shows
battery discharging time is from 17:00 to 22:00 h. Charging and discharging of
battery, in both cases in single hour is 1.5 kWh [9], whereas efficiency of battery
charging and discharging is taken as 7.2 cents/kWh [15]. Figure 6 shows mape
values of 3 different forecasting algorithms including BP, SVM, SAE. Historical
data has been used for forecasting solar power that is called step 1, after retrieval
of foretasted data, it is added in historic data and used for prediction that is
termed as step 2 and so on. It is clearly shown in Fig. 6 that mape increases
after each iteration. Simulation shows mape acquired as a result of forecasting
techniques implementation. SAE has performed best among all, therefore error
of SAE is taken as prediction error in objective function of microgrid. There
exist two scenario, that prediction error can be positive, i.e., Δ > 0 and it can
be negative, i.e., Δ < 0 respectively. Δ > 0 shows that predicted output of solar
power is less than the actual. Which leads microgrid to purchase it from utility.
In the second scenario, Δ < 0 is predicted value is greater than actual value.
Which shows microgrid will not procure electricity from utility [12].
5 Conclusion
In this paper, energy management system is focused mainly, which includes
single microgrid and multiple users. To use renewable resources generated energy
optimally, solar power generation forecasting is performed in which three existing
techniques are used and comparison among them is performed. The forecasting
error is, thus brought into account for analyzing the impact on pay off of both
stakeholders in game theory. Stackleberg game theory is utilized in this paper
to prove nash equilibrium among two player, i.e., user and microgrid. In future
work, cooperative energy management will be emphasized.
References
1. Erol-Kantarci, M., Mouftah, H.T.: Smart grid forensic science: applications, chal-
lenges, and open issues. IEEE Commun. Mag. 51(1), 68–74 (2013)
2. Saad, W., Han, Z., Poor, H.V., Basar, T.: Game-theoretic methods for the smart
grid: an overview of microgrid systems, demand-side management, and smart grid
communications. IEEE Sig. Process. Mag. 29(5), 86–105 (2012)
3. Ipakchi, A., Albuyeh, F.: Grid of the future. IEEE Power Energ. Mag. 7(2), 52–62
(2009)
4. Liang, X., Li, X., Lu, R., Lin, X., Shen, X.: UDP: usage-based dynamic pricing
with privacy preservation for smart grid. IEEE Trans. Smart Grid 4(1), 141–150
(2013)
5. Sinha, A., Malo, P., Frantsev, A., Deb, K.: Finding optimal strategies in a multi-
period multi-leader follower Stackelberg game using an evolutionary algorithm.
Comput. Oper. Res. 41, 374–385 (2014)
6. Bingtuan, G.A.O., Xiaofeng, L.I.U., Cheng, W.U., Yi, T.A.N.G.: Game-theoretic
energy management with storage capacity optimization in the smart grids. J. Mod.
Power Syst. Clean Energ. 1–12 (2018)
1106 A. Naz et al.
7. Kanchev, H., Lu, D., Colas, F.: Energy management and operational planning of
a microgrid with a PV-based active generator for smart grid applications. IEEE
Trans. Industr. Electron. 58(10), 4583–4592 (2011)
8. Tazvinga, H., Xia, X.H.: Minimum cost solution of photovoltaic-diesel-battery
hybrid power systems for remote consumers. Sol Energ. 96, 292–299 (2013)
9. http://www.elia.be/en/grid-data/power-generation/Solar-power-generation-
data/Graph
10. Darivianakis, G., Georghiou, A., Smith, R.S., Lygeros, J.: A stochastic optimiza-
tion approach to cooperative building energy management via an energy hub. In:
2015 IEEE 54th Annual Conference on Decision and Control (CDC), pp. 7814–
7819. IEEE, December 2015
11. Valencia, F., Collado, J., Sáez, D., Marı́n, L.G.: Robust energy management system
for a microgrid based on a fuzzy prediction interval model. IEEE Trans. Smart Grid
7(3), 1486–1494 (2016)
12. Zhou, Z., Wang, Y., Wu, Q.J., Yang, C.N., Sun, X.: Effective and efficient global
context verification for image copy detection. IEEE Trans. Inf. Forensic Secur.
12(1), 48–63 (2017)
13. Zhou, Z., Xiong, F., Huang, B., Xu, C., Jiao, R., Liao, B., Yin, Z., Li, J.: Game-
theoretical energy management for energy internet with big data-based renewable
power forecasting. IEEE Access 5, 5731–5746 (2017)
14. Khalid, R., Javaid, N., Rahim, M.H., Aslam, S., Sher, A.: Fuzzy energy manage-
ment controller and scheduler for smart homes. Sustain. Comput. Inf. Syst. 21,
103–118 (2019)
15. Ahmad, A., Javaid, N., Mateen, A., Awais, M., Khan, Z.: Short-term load forecast-
ing in smart grids: an intelligent modular approach. Energies 12(1), 164 (2019)
16. Naz, M., Iqbal, Z., Javaid, N., Khan, Z.A., Abdul, W., Almogren, A., Alamri,
A.: Efficient power scheduling in smart homes using hybrid Grey Wolf differential
evolution optimization technique with real time and critical peak pricing schemes.
Energies 11(2), 384 (2018)
17. Nadeem, Z., Javaid, N., Malik, A.W., Iqbal, S.: Scheduling appliances with GA,
TLBO, FA, OSR and their hybrids using chance constrained optimization for smart
homes. Energies 11(4), 888 (2018)
18. Sher, A., Khan, A., Javaid, N., Ahmed, S., Aalsalem, M., Khan, W.: Void hole
avoidance for reliable data delivery in IoT enabled underwater wireless sensor net-
works. Sensors 18(10), 3271 (2018)
19. Mossoud, S., Wollenberg, B.: Toward a smart grid: power delivery for the 21st
century. IEEE Power Energ. Mag. 3, 34–41 (2005)
20. Gungor, V.C., Lu, B., Hancke, G.P.: Opportunities and challenges of wireless sensor
networks in smart grid. IEEE Trans. Ind. Electron. 57(10), 3557–3564 (2010)
21. Wang, Y., Mao, S., Nelms, R.M.: Distributed online algorithm for optimal real-
time energy distribution in the smart grid. IEEE Internet Things J. 1(1), 70–80
(2014)
22. Muralitharan, K., Sakthivel, R., Shi, Y.: Multiobjective optimization technique for
demand side management with load balancing approach in smart grid. Neurocom-
puting 177, 110–119 (2016)
23. Yin, S., Gao, X., Karimi, H.R., Zhu, X.: Study on support vector machine-based
fault detection in Tennessee Eastman process. Abstr. Appl. Anal. 2014, 8 (2014)
24. Dai, Y.M., Gao, Y.: Real-time pricing strategy with multi-retailers based on
demand-side management for the smart grid. Proc. Chin. Soc. Electric. Eng.
34(25), 4244–4249 (2014)
Electricity Load Forecasting for Each Day
of Week Using Deep CNN
Sajjad Khan1 , Nadeem Javaid1(B) , Annas Chand2 ,

Abdul Basit Majeed Khan3 , Fahad Rashid4 , and Imran Uddin Afridi1
1
2
COMSATS University Islamabad, Abbottabad Campus,
Abbottabad 22010, Pakistan
3
Abasyin University Islamabad, Islamabad 44000, Pakistan
4
Bahria University Islamabad, Islamabad 44000, Pakistan
Abstract. In smart grid, precise and accurate electricity load forecast-

ing is one of the most challenging tasks. It is due to the high volatile, non-
stationary and non-linear behavior of electricity load data. In this paper,
a Deep Convolution Neural Network (DCNN) model is proposed to fore-
cast the electricity load for each day of the week of Victoria (Australia).
To forecast the electricity load for one day of the week, we analyzed the
electricity load data consumed on the same day for the previous three
months. To show the usefulness of our proposed scheme, comparison is
made with the state of the art forecasting models namely recurrent neural
network, extreme learning machine, CNN and auto regressive integrated
moving average. Results show that the proposed DCNN has the low-
est mean absolute percentage error, mean absolute error and root mean
square error of 2.1%, 138.771 and 116.417.
Keywords: Smart grid · Forecasting · Energy management ·

Neural Network · Deep learning
1 Background
Forecasting the accurate electricity load demand plays a key role in demand
side management. It helps the generation companies to develop a tolerable pro-
duction plan and avoid energy wastage. For electricity load forecasting many
techniques have been used in literature. These prediction techniques are mainly
classified into three main groups, i.e., data driven techniques, classical techniques
and Artificial Intelligence (AI) techniques. Data driven techniques predict the
desired outcomes on the basis of analysing historical data. Classical techniques
comprises of the statistical and mathematical methods like Autoregressive Inte-
grated Moving Average (ARIMA), Seasonal ARIMA and random forest etc.

https://doi.org/10.1007/978-3-030-15035-8_107
1108 S. Khan et al.
Such methods are effective for one step forecast. AI methods inhibit the behavior
of biological neurons e.g., Feed Forward Neural Network (FFNN), Convolution
Neural Network (CNN) and Long Short Term Memory (LSTM) etc.
In [2], Bayesian network and Multi Layer Perceptron (MLP) are used for
short and long term load forecasting. Bayesian network performs behavioral
analytics using the data obtained from the smart meters. Number of experi-
ments were performed on the obtained data. For Short Term Load Forecasting
(STLF), multiple linear regression is used in [3]. However, it has the limitation
that it can not be used for long term prediction. The authors in [4] performed
electricity load forecasting on the basis of weather data obtained from the mete-
orological department. In this work, load prediction is performed using residual
Neural Network (NN). A framework of different models is used for prediction
in [5]. This framework consists of Gated Recurrent Unit (GRU), Recurrent Neu-
ral Network (RNN) and LSTM. GRU is a variant of the RNN. It outperformed
LSTM and many other statistical models in terms of prediction accuracy. Based
on the simulation results, it is proved that deep learning models are better than
the statistical models. The authors in [6] used Restricted Boltzmann machine
and Rectified Linear Unit (ReLU) for training the data and electricity load
prediction respectively. Discrete wavelet transform and inconsistency rate meth-
ods are used for optimal feature selection from the features set in [7]. It helps
in dimensionality reduction. The parameter tuning of Support Vector Machine
(SVM) is performed using Sperm whale algorithm. Authors in [8] proposed a
model for STLF. In this work, for feature selection, Mutual Information (MI) is
used whereas, better prediction results are achieved by modifying the Artificial
Neural Network (ANN). Extreme Learning Machine (ELM) is used in [9] for
efficient load prediction.
The authors in [10] proposed an intelligent approach to forecast electricity
load using MI and ANN. In [11] the authors forecast electricity load and price
using enhanced CNN and enhanced SVM. ELM is a generalized single hidden
layer FFNN learning algorithm. It is proved to be effective for regression and
classification purposes. Authors used the Neural Network (NN) for achieving
better load prediction in [12]. Input weights and biases are randomly assigned
in ELM learning process, whereas output weights are calculated using Moore-
Penrose generalized inverse technique. Efficient load forecasting is performed
using sparse Bayesian ELM in [13]. This method is used for multi-classification
purpose. Authors in [14] used Particle Swarm Optimization (PSO) and discrete
PSO techniques for predicting day ahead electricity load. The Grey wolf opti-
mizer is used in combination with NN to optimize weights and biases in [15].
The simulation results proved that weights and biases optimization technique
improved the prediction accuracy. Back propagation technique is used for train-
ing ELM in [16]. Training is done using context neurons as input to hidden and
input layer of ELM. Authors in [17] enhanced RNN to achieve better prediction
accuracy by adjusting weights using previous iteration errors. Moreover, adjust-
ing biases and neurons selection also affects the prediction accuracy. The authors
in [18] developed a knowledge based system for STLF. In [19], 24 h ahead predic-
tion of cooling load of building is performed using deep learning. The simulation
Electricity Load Forecasting for Each Day of Week Using Deep CNN 1109
results showed that deep learning techniques improved the cooling load predic-
tion. Similarly in [20], RNN is used for electricity load prediction. In this work,
RNN groups the consumers into pool of inputs. It then handles the consumers
data as aggregated data, which reduces the computational time and resource
utilization.
1.1 Motivation and Problem Statement

The authors in [2] and [3] forecast electricity load using shallow NN by analysing
the historic load data. In [4], the authors used weather data to forecast the
electricity load demand using RNN. ELM is used to forecast the electricity load
demand in [16]. In [19], the cooling load of a residential building is forecasted
using a deep NN. The aforementioned work motivates us to forecast electricity
load using deep NN.
MLP is used in [2] and [3] for short term load forecasting. Similarity a single
layer FFNN is used in [4] to forecast load. In [16], a back propagation technique
is used. Similarly, the authors in [19] used a deep RNN to forecast the load
demand of residential buildings. None of the above mentioned work used Deep
CNN (DCNN) to forecast electricity load demand for all days of the week.
1.2 Contribution
The main contribution of this paper are following:
• An DCNN is proposed to forecast the electricity load of Victoria.
• A performance comparison of the benchmark schemes with the proposed
scheme is performed.
The rest of this paper is organized as follows. Section 2 is the proposed sys-
tem model. The proposed and existing schemes used to forecast electricity load
are discussed in Sect. 3. A detailed discussion on the simulation results of each
scheme is presented in Sect. 4. Section 5 concludes the paper along with the
future directions of our work.
2 System Model
Figure 3 presents a complete picture of our proposed system model. Forecasting
electricity load belongs to the paradigm of Time Series (TS) forecasting. Accu-
rate load forecasting using TS data is not an easy task. Therefore, in this paper,
we have first converted the TS electricity load data into a supervised learning
problem. Afterwards, training and testing is performed to forecast the electricity
load using the proposed model.
To convert the TS electricity load data into a supervised learning problem, we
have first extracted all days of the week separately from the historical load data.
All Mondays are extracted to form a Monday group. Similarly, all Tuesdays are
extracted to form a Tuesday group etc. This processes is repeated for all days.
1110 S. Khan et al.
After the formation of seven groups, we further divide each group into a train
and test dataset sub-group. Every train data sub-group contains load data of
eight days. Whereas, the test dataset sub-group contains the load data of the
ninth day. After the formation of train and test data sub-groups, these groups
are appended one after the other in such a way that all Mondays are followed by
all Tuesdays, Similarly all Tuesdays are followed by Wednesdays etc. Once all
the days are grouped into one set of train and test dataset, we create the input
and output variables such that output to Xi is Xi+1 and output of Xn is X1 .
The process of input and output variable formation is shown in Fig. 1.
Monday t1 t2 ... n-1 n t1

Input variables
Tuesday t1 t2 ... n-1 n t1
Output
Sunday t1 t2 ... n-1 n t1
variables
Fig. 1. Input and output variables formation
2.1 Dataset Description
In this paper, the dataset released by the Australian energy market operator is
used to forecast the electricity load of Victoria. The dataset is publicly available
at [1]. In this paper, the dataset used for forecasting contains load data from the
last week of April 2018 till the second week of July 2018. The load values are
recorded after every thirty minutes in a day and there are forty eight observations
in one day. Figure 2 shows the electricity load data used in this paper.
Electricity load (MW)
7500
5500
3500
0 500 1000 1500 2000 2500 3000
Time (30 min interval)
Fig. 2. Electricity load of Victroia

2.2 Performance Metrics
In forecasting the electricity load, researchers have used a number of performance

metric parameters. However, none of these performance evaluation metrics have
been marked as the standard parameters for evaluating the performance of a fore-
casting model. In order to evaluate, the effectiveness of our proposed scheme, we
used three performance metric parameters. These parameters are Mean Abso-
lute Percentage Error (MAPE), Mean Absolute Error (MAE) and Root Mean
Square Error (RMSE). Equations 1, 2 and 3 presents the mathematical formula
of MAPE, MAE and RMSE respectively. Furthermore, to assess the performance
of each forecasting model, the lower the value of each performance metric metric
indicator, the higher the forecasting accuracy will be.
N
1 Xi − Yi
M AP E = ( | |) × 100 (1)
N t=1 Yi
N
1
M AE = |Xi − Yi |2 (2)
N t=1

1 N
RM SE = |Xi − Yi |2 (3)
N t=1
Here Xi presents the actual load value, whereas, Yi presents the load forecast
values. To evaluate the forecasting results of a model, the lower the performance
metric parameter values, the better the results would be.
3 Existing and Proposed Forecasting Schemes
In this section we discussed the proposed and existing benchmark schemes used
to forecast the electricity load.
3.1 RNN
Among all the NN models, LSTM [21] is the most common type of RNN due
to its ability of memorizing information for a longer period of time. In LSTM,
every node act as a memory cell to store information. Furthermore, each node
has its own activation function. In order to process the input sequence, RNN
use the internal states. LSTM also have internal states, however, the difference
between other NN and LSTM is that in LSTM the old state becomes the input
to the next state along with the current input.
1112 S. Khan et al.
Start
Pre processing
Days extraction and dataset formation

Pre processing
Deep CNN structure
Selection of loss function
Selection of an activation function

Add hidden layers
Training and testing

Parameter
tuning Electricity load forecasting
Performance metrics
MAPE
Yes/No
MAE
RMSE
End
3.2 CNN
The most successful type of neural network to solve various problems in medical
imaging and computer vision is CNN [25]. It is due to the fact that CNN can
handle large amount of input sequences. Furthermore, it has the ability to extract
the hidden features in the input data. The architecture of CNN comprises of
convolution, fully connected, dense, flatten and max pooling layers. Commonly,
the deeper the CNN model, the higher the accuracy will be. In this paper, CNN
has one convolution, one dropout and two dense layers.
3.3 ARIMA
It is one of the most commonly used statistical technique for TS forecasting.
ARIMA as a TS forecasting model is proposed by Box and Jenkins [22] as a
combination of Auto Regression (AR) and Moving Average (MA). The key idea
to combine AR and MA is to approximate the real world representation in a
more accurate and precise way. The main advantage of using ARIMA to forecast
electricity load is that it can accurately forecast the future load values if the TS
is a stationary series. However, if the TS is not stationary, then the input series
is converted to a stationary TS by taking the first or second difference.
3.4 ELM
Huang et al. in [23] and [24] proposed ELM, which a type of feed forward neu-
ral network. Due to its generalization and learning abilities, ELM as forecast-
ing model has gained popularity in several fields. The core advantage of ELM
includes: (i) parameters of the hidden layers do not require fine tunning: (ii)
biasness and the input weights to ELM are randomly assigned.
3.5 DCNN
We enhance the forecasting performance of CNN by adding extra layers. Increas-
ing the number of layers in CNN scale up the size and complexity of the proposed
DCNN. However, in this work, our focus is to minimize the difference between
the actual and forecast values. The proposed DCNN comprises of convolution
and dense layers. A dropout layer at 20% along with ReLU activation function.

This section discusses the forecasting results. To affirm the usefulness of our
scheme, the proposed DCNN is compared with four benchmark schemes. To
better understand the forecasting behavior of each model, this section is further
divided in to five subsections.
4.1 Forecasting Results of RNN

The electricity load forecasted by RNN for Victoria is shown in Fig. 7. In this
figure, it can be clearly seen that there is a huge difference between the actual
and forecasted electricity load using RNN. From the performance metric graphs
shown in Figs. 4, 5 and 6 it is concluded that RNN has the worst performance
in forecasting electricity load as compared to all the benchmark schemes used in
this paper. According to the performance metric values of RNN, Thursday has
the lowest error values. The minimum MAPE, MAE and RMSE values using
RNN to forecast electricity load of Victoria are 13.439%, 642.835 and 620.471
respectively. Furthermore, Monday has the worst performance metric values as
compare to all days of the week. To further study the usefulness of RNN in
forecasting electricity load of Victoria, the performance metric values for all
days of the week are displayed in Table 1.
1114 S. Khan et al.
14 Mon Thu Sat

Tue Fri Sun
12 Wed
10
MAPE Scores
8
0
ELM RNN DCNN CNN ARIMA
Fig. 4. MAPE of the five schemes using Victoria dataset
4.2 Forecasting Results of CNN
Figure 7 shows the electricity load forecast of Victoria for all days using CNN.
As shown in this figure, the performance of CNN is slightly better as compare to
RNN. However, the difference in the forecasted and actual electricity is still very
large. The bar plots in Figs. 4, 5 and 6 presents the performance errors metrics
of CNN. According to Table 1, CNN has the lowest error values in forecasting
the load of Thursday and Saturday. The MAPE, MAE and RMSE values for
CNN are 7.20%, 385.869 and 359.788 respectively. Whereas, in forecasting the
electricity load of Monday and Friday CNN has the worst performance metric
values.
700 Mon Thu Sat

Tue Fri Sun
600 Wed
500
MAE Scores
400
300
200
100
0
Fig. 5. MAE of the five schemes using Victoria dataset

4.3 Forecasting Results of ARIMA

According to Fig. 7, ARIMA achieved average results in forecasting electricity
load for Victoria as compare to CNN and RNN. Figure 7 shows that at some time
slots the load forecasted by ARIMA is exactly the same as actual load. However,
at some time slots the load forecasted by ARIMA is totally unacceptable e.g., the
first time slot of Friday and Tuesday. According to the performance metric graphs
in Figs. 4, 5 and 6 the performance of ARIMA in forecasting load for Victoria
is better on Wednesday as compare to other days of the week. The minimum
MAPE, MAE and RMSE achieved in forecasting electricity load by ARIMA for
Victoria is 3.58%, 238.154 and 194.830 respectively. Similarly, ARIMA has the
worst performance metric values in forecasting load of Tuesday.
Mon Thu Sat

700
Tue Fri Sun
Wed
600
500
RMSE Scores
400
300
200
100
0
Fig. 6. RMSE of the five schemes using Victoria dataset
4.4 Forecasting Results of ELM

Table 1 presents the performance metric values of ELM for all days of the week.
As shown in Fig. 7, ELM is the second best choice to forecast electricity load
for Victoria. From the performance metric graphs in Figs. 4, 5 and 6 Wednesday
has the least error values. The MAPE, MAE and RMSE values for Wednesday
are 2.170%, 342.677 and 259.769 respectively. Similarly, ELM has the worst
performance metric values in forecasting electricity load of Monday.
4.5 Forecasting Results of DCNN

The performance metric graphs in Figs. 4, 5 and 6 shows that the proposed
DCNN has the lowest performance metric values for all days of the week as
compare to all the benchmark schemes. According to Fig. 7, the load forecasted
by the proposed DCNN for Victoria almost overlaps the actual values for all
days of the week. Table 1 shows the performance metric values for all days of the
week using Deep CNN. From this table, it can be clearly seen that the minimum
1116 S. Khan et al.
7500
Monday
5500
3500
7500
Tuesday
5500
3500
7500
Wednesday
5500
Elect ricit y load (MW)
3500
7500
Thursday
5500
3500
7500
Friday
5500
3500
7500
Sat urday
5500
3500
7500
Sunday
5500
3500
0 4 8 12 16 20 24
DCNN RNN Hour(s)

ARIMA ELM
CNN Actual
Fig. 7. Forecasting results of all schemes for Victoria

Table 1. Performance metric values

MAPE
Monday 4.303 13.969 2.893 7.515 4.629
Tuesday 3.577 13.708 2.805 7.263 5.047
Wednesday 2.170 13.781 2.152 7.379 3.583
Thursday 2.606 13.439 2.489 7.370 3.984
Friday 3.846 13.937 2.705 7.609 4.633
Saturday 4.270 13.637 2.924 7.200 4.538
Sunday 4.017 13.787 2.927 7.370 4.578
MAE
Monday 342.677 767.183 218.879 469.614 347.052
Tuesday 271.484 732.568 201.115 441.339 352.144
Wednesday 152.037 683.748 138.771 401.976 238.154
Thursday 194.463 642.835 184.972 385.869 265.644
Friday 288.611 738.617 192.779 449.356 343.053
Saturday 336.182 752.452 218.885 455.808 351.269
Sunday 306.203 745.615 210.575 453.192 332.788
RMSE
Monday 259.769 732.158 171.732 415.240 273.986
Tuesday 209.630 701.686 162.286 392.409 293.096
Wednesday 117.390 666.721 116.417 376.971 194.830
Thursday 139.007 620.471 132.354 359.788 212.860
Friday 220.306 705.959 148.614 404.669 256.654
Saturday 256.228 716.093 172.620 399.010 266.153
Sunday 235.286 711.188 167.861 400.558 261.256
MAPE, MAE and RMSE in forecasting electricity load for Victoria are 2.15%,
138.771 and 116.417 respectively. Moreover, the lowest error values in forecasting
load using DCNN are achieved for Thursday. Similarly, the highest MAPE, MAE
and RMSE are 2.927%, 218.885 and 172.620. To sum up, by comparing the
usefulness of our proposed DCNN with the benchmark schemes, it can be said
that the proposed scheme has the best result in forecasting the electricity load
for all days.
5 Conclusion and Future Work

For utilities and generation companies, forecasting the actual demand of electric-
ity minimizes the power risk and production cost of electricity. NN are gaining
popularity in forecasting the electricity load demand because of their ability
1118 S. Khan et al.
to handle non linear and non stationary structure of the electricity load data.
In this paper, a DCNN is proposed to forecast the electricity load demand for
each day of week. Our experimental results illustrate that the proposed scheme
has the lowest MAPE, MAE and RMSE scores as compare to the benchmark
schemes. The proposed DCNN exceeds the existing benchmark schemes by scor-
ing the lowest MAPE, MAE and RMSE of 2.152%, 138.771 116.417. ELM has
the second best performance metric scores of 2.17, 152.037 and 117.390. ARIMA
achieved average results whereas, CNN and RNN has the worst performance
metric values. In future, we plan to investigate the impact of feature engineering
in forecasting the electricity load and price. Furthermore, to validate the effec-
tiveness of our proposed scheme, we plan to forecast the electricity load of other
states of Australia such as New South Wales and Queensland etc.
References
1. https://www.aemo.com.au/. Accessed 01 Jan 2019
and energy consumption forecasting. Energies 11(2), 452 (2018)
3. Ahmad, T., Chen, H.: Short and medium-term forecasting of cooling and heating
load demand in building environment with data-mining based approaches. Energy
Build. 166, 460–476 (2018)
4. Chen, K., Chen, K., Wang, Q., He, Z., Hu, J., He, J.: Short-term load forecasting
with deep residual networks. IEEE Trans. Smart Grid (2018)
5. Jesus, L., De Ridder, F., De Schutter, B.: Forecasting spot electricity prices: Deep
Energy 221, 386–405 (2018)
6. Ryu, S., Noh, J., Kim, H.: Deep neural network based demand side short term load
forecasting. Energies 10(1), 3 (2016)
7. Liu, J., Li, C.: The short-term power load forecasting based on sperm whale algo-
rithm and wavelet least square support vector machine with DWT-IR for feature
selection. Sustainability 9(7), 1188 (2017)
8. Ashfaq, A., Javaid, N., Guizani, M., Alrajeh, N., Khan, Z.A.: An accurate and fast
converging short-term load forecasting model for industrial applications in a smart
grid. IEEE Trans. Ind. Inform. 13(5), 2587–2596 (2017)
9. Huang, G.-B., Zhou, H., Ding, X., Zhang, R.: Extreme learning machine for regres-
sion and multiclass classification. IEEE Trans. Syst. Man Cybern. Part B (Cybern.)
42(2), 513–529 (2012)
11. Zahid, M., Ahmed, F., Javaid, N., Abbasi, R.A., Kazmi, H.S.Z., Javaid, A., Bilal,
M., Akbar, M., Ilahi, M.: Electricity price and load forecasting using enhanced
convolutional neural network and enhanced support vector regression in smart
grids. Electronics 8(2), 122 (2019)
12. Fallah, S., Deo, R., Shojafar, M., Conti, M., Shamshirband, S.: Computational
intelligence approaches for energy load forecasting in smart energy management
grids: state of the art, future challenges, and research directions. Energies 11(3),
596 (2018)
13. Luo, J., Vong, C.-M., Wong, P.-K.: Sparse Bayesian extreme learning machine for
multi-classification. IEEE Trans. Neural Netw. Learn. Syst. 25(4), 836–843 (2014)
14. Yu, J., Wang, S., Xi, L.: Evolving artificial neural networks using an improved PSO
and DPSO. Neurocomputing 71(4–6), 1054–1060 (2008)
15. Zeng, Y.-R., Zeng, Y., Choi, B., Wang, L.: Multifactor-influenced energy consump-
tion forecasting using enhanced back-propagation neural network. Energy 127,
381–396 (2017)
16. Saremi, S., Mirjalili, S.Z., Mirjalili, S.M.: Evolutionary population dynamics and
grey wolf optimizer. Neural Comput. Appl. 26(5), 1257–1263 (2015)
17. Ertugrul, Ö.F.: Forecasting electricity load by a novel recurrent extreme learning
machines approach. Int. J. Electr. Power Energy Syst. 78, 429–435 (2016)
18. Khan, M., Javaid, N., Naseem, A., Ahmed, S., Riaz, M., Akbar, M., Ilahi, M.:
Game theoretical demand response management and short-term load forecasting
by knowledge based systems on the basis of priority index. Electronics 7(12), 431
(2018)
19. Fan, C., Xiao, F., Zhao, Y.: A short-term building cooling load prediction method
using deep learning algorithms. Appl. Energy 195, 222–233 (2017)
20. Shi, H., Minghao, X., Li, R.: Deep learning for household load forecasting-a novel
pooling deep RNN. IEEE Trans. Smart Grid 9(5), 5271–5280 (2018)
21. Hochreiter, S., Schmidhuber, J.: Long short-term memory. Neural Comput. 9(8),
1735–1780 (1997)
22. Box, G.E.P., Jenkins, G.M., Reinsel, G.C., Ljung, G.M.: Time Series Analysis:
Forecasting and Control. Wiley, Hoboken (2015)
23. Huang, G.-B., Zhu, Q.-Y., Siew, C.-K.: Extreme learning machine: a new learning
scheme of feedforward neural networks. In: 2004 Proceedings of the IEEE Interna-
tional Joint Conference on Neural Networks, vol. 2, pp. 985–990. IEEE (2004)
24. Huang, G.-B., Chen, L., Siew, C.K.: Universal approximation using incremental
constructive feedforward networks with random hidden nodes. IEEE Trans. Neural
Netw. 17(4), 879–892 (2006)
25. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep con-
volutional neural networks. In: Advances in Neural Information Processing Sys-
tems, pp. 1097–1105 (2012)
Short Term Load Forecasting
Using XGBoost
Raza Abid Abbasi1 , Nadeem Javaid1(B) , Muhammad Nauman Javid Ghuman2 ,

Zahoor Ali Khan3 , Shujat Ur Rehman2 , and Amanullah2
1
COMSATS Institute of Information Technology, Islamabad 44000, Pakistan
2
Quaid-i-Azam University, Islamabad 44000, Pakistan
3
Computer Information Science, Higher Colleges of Technology,
Fujairah 4114, UAE
http://www.njavaid.com/
Abstract. For efficient use of smart grid, exact prediction about the in-
future coming load is of great importance to the utility. In this proposed
scheme initially we converted daily Australian energy market operator
load data to weekly data time series. Furthermore, we used eXtreme Gra-
dient Boosting (XGBoost) for extracting features from the data. After
feature selection we used XGBoost for the purpose of forecasting the elec-
tricity load for single time lag. XGBoost perform extremely well for time
series prediction with efficient computing time and memmory resources
usage. Our proposed scheme outperformed other schemes for mean aver-
age percentage error metric.
1 Introduction
Energy production and consumption difference minimization is a challenging
task these days. Efficient consumption of electricity is one good solution to this
problem. Researchers has done a lot of work in for introducing cost effective
and efficient energy utilization systems. Next generation Smart Grid (SG) is
the most attractive solution so far. SG is the integration of information and
communication technology in traditional grid which makes it intelligent power
grid supporting real-time information exchange between producer and consumer.
SG enables the energy efficiency optimization. More precisely, the SG needs an
accurate forecasting of the energy load for more productive application.
AN increasing attempt of deregulating strength markets to shape a more
dependable, green, and price-effective system with the aid of improving com-
petitions has been witnessed in world’s important economies [1,2]. In the lib-
eralized markets, the strength is commoditized and consequently its price is
dynamic. Due to the fee variant, pricing power correctly will become important
to generate profits, schedule strength productions, and plan load responses [3–
7]. The accurate power charge forecasting is helpful to decide the strength rate
and accordingly is precious. As liberalized electricity markets include types, day-
ahead and real-time [8], it’s miles meaningful to discuss both of the day-ahead
https://doi.org/10.1007/978-3-030-15035-8_108
STLF Using XGBoost 1121
and online forecasting of the power fee. The energy rate forecasting has been
vigorously studied inside the literature. From the application aspect, the fore-
casting of the energy rate in unique deregulated markets of essential economies
around the world has been stated [9–15].
The Rest of the paper is structured as defined next. The mostly used practi-
cle techniques for load forecasting are discussed in Sect. 2. The Sect. 3 enlightens
the different problems associated with the load prediction. The proposed model
for load prediction and the evaluation metrices are explained in Sect. 4. Exper-
imental results are depicted and highlighted in Sect. 5. Conclusion about the
work done in this paper is expressed in Sect. 6 at the end.
2 Related Work
Authors in [16] proposed a novel practical methodology using quantile regression

mean on a set of sisters point forecasts. Data from GEFCom2014 probabilistic
load track was used to forecast for developing probabilistic load forecasts. The
suggested scheme has dual advantages, where It will strength the advancement in
the field of point load forecasting, it is not dependent on the high quality expert
predictions. Scheme proposed in this work produces better results as compared to
the benchmark methods. “Recency effect” a psychology term is used by authors
in [17]. They exploited the fact that power consumption demand is influenced by
the temperature of the earlier hours. Authors produced a ample study to show
the effect of recency with the help of big data. Modern computing power is used
in order to decide how many lagged temperature are required for catching the
recency effect completely without affecting the predicting accuracy.
In [18] authors proposed a scheme for big data analytics in smart grid which
aim at reducing the electricity cost for users. Moreover they explored the individ-
ual components needed for an improved decision support system for the purpose
of energy saving. The presented framework has four different layers in its archi-
tecture, i.e., smart grid, data accumulation, an analysis counter and supporting
web portal. Future power consumption is fore-casted and optimized through a
innovative composite nature inspired meta heuristic prediction scheme. A versa-
tile optimization algorithm works as a backbone for the analytics counter that
helps in achieving accurate results. The proposed novel framework is the major
contribution of this work, which supports the energy saving decision process.
This contribution is the basis for full scale, Smart Decision Support System
(SDSS). SDSS can identify the usage pattern of an individual user which helps
in enhancing the efficiency of energy usage where improving the accuracy of fore-
casted energy demands. Authors in [19] used forecasting analytics while focusing
the extraction of related external features. More explicitly the proposed scheme
predicts the spot prices in German energy market in relation to the historical
data of prices and weather features. Least Absolute Shrinkage Selection Opera-
tion (LASSO) finds the related weather stations where implicit variable selection
is executed by Random Forest (RF). This work enhanced the prediction accuracy
with respect to Mean Average Error (MAE) by 16.9%.
A novel modeling scheme for electricity price prediction is introduced in [20].

Four different deep learning models are suggested by Lago et al. for forecast-
ing electricity prices that lead to advancement in forecasting accuracy. Authors
proposed despite the presence of a good number of electricity price forecast-
ing methods, still a benchmark is missing. This work compared and evaluated
27 different common techniques used for electricity price prediction and then
proved how the proposed models outperform the state of the art techniques
those are significant. Wang et al. used Stacked De-noising Auto-encoder and
Random Samples RS-SDA for live and next day hourly price prediction. In [21]
short term forecasting of the electricity price is performed using data driven
scheme. Deep Neural Networks type, SDA and its extended version RS-SDA are
used to forecast the electricity price hourly for the data collected from different
states of United States. This research is focused on next day hourly prediction
and the live hourly prediction. SDA defined models are assessed in comparison
with conventional neural network and support vector machine, where next day
prediction SDA models accuracy is assessed in comparison industrial model.
In [22] Lagoa et al. introduced two distinct schemes for combining market
incorporation in energy price prediction and to enhance the forecasting depic-
tion. First scheme suggested a DNN that examines features from linked markets
to enhance the forecasting results in a community market. Features importance
is calculated using a innovative feature selection scheme that contains the opti-
mization and functional analysis of variance. Second scheme forecasts the prices
from two adjacent markets simultaneously which bring the accuracy metric Sym-
metric Mean Absolute Error (SMAPE) even further lower. Raviv et al. worked
on predicting next day energy prices while utilizing hourly prices in [23]. This
work exhibit that dismantled hourly rates include handy forecasting facts for the
daily typical prices in the Nord pool market. It is evaluated that the multivari-
ate patterns for the complete group of hourly prices considerably go better than
univariate patterns of the daily normal price. Multivariate models reduce RMSE
upto 16%. In [24] authors worked on electrical load forecasting on the basis of
pre analysis and weight coefficients optimization. A novel scheme is introduced
exploiting the features of electrical load data i.e., capacity to effectively calcu-
late the seasonality and nonlinearity. The proposed new scheme can use up the
advantages stay away from disadvantages of the individual schemes. In suggested
combined scheme the data fore analyzation is adapted so that conflicts can be
minimized in the data, where weight factors are adjusted using cuckoo search
in the combined model. The newly proposed scheme outperforms the individual
forecasting models regarding forecast performance.
Singh et al. worked on the amount of power consumed prediction in [25].
An intellectual data mining scheme is proposed that can evaluate, predict and
reflect electricity time series to disclose numerous temporary energy using pat-
terns. These patterns help to identify appliance usage relationship with time
i.e., hour of day, week, month e.t.c, and appliance usage relationship with other
appliances. This identification basis for the understanding the customer usage
behavior, energy load forecasting and the price forecasting. Authors proposed
Bayesian network forecasting, constant analysis of data through data mining and
unsupervised data accumulating for electricity consumption prediction. In [26]
authors proposed a short-lived electricity load prediction scheme for academic
buildings. This work used 2-stage forecasting analysis for the productive working
of their energy system. Energy consumption data is collected from different uni-
versities and moving average method is used for finding the energy load pattern
according to week day. Random Forest (RF) technique is used for forecasting
the daily energy load. RF performance is assessed using cross-validation on time
series.
González et al. predicted electricity price adopting functional time series
using a New Hilbertian ARMAX model in [27]. Suggested scheme has a lin-
ear regression structure, where functional variables are operated by functional
parameters. Where functional parameters are fundamental entities with linearly
combined kernels as sigmoid operations. Quasi-Newton model is used for param-
eters optimization in sigmoid which minimizes the sum of squared error. Data
integrity attacks affect the results of load prediction models i.e., artificial neural
network, multiple linear regression, support vector regression and fuzzy interac-
tion regression). Authors in [28] worked on exposing the consequences of these
attacks. We begin by simulating some knowledge integrity attacks through the
random injection of some multipliers that follow a traditional or uniform distri-
bution into the load series. Then, the four same load prognostication models are
used to generate one-year-ahead ex post purpose forecasts so as to supply a com-
parison of their forecast errors. The results show that the support vector regres-
sion model is most robust, followed closely by the multiple rectilinear regression
model, whereas the fuzzy interaction regression model is that the least sturdy of
the four. withal, all four models fail to supply satisfying forecasts once the size
of the info integrity attacks becomes giant. This presents a serious challenge to
each load forecasters and therefore the broader prognostication community: the
generation of correct forecasts beneath knowledge integrity attacks.
Dong et al. worked on the energy management in a microgrid. Bayesian-
optimization-algorithm (BOA) is used for a single SG using house. Authors
in [29] articulates the enhancement beyond the closed form equitable function
equation, and work out on it using BOA based data-driven technique. We can
consider the suggested technique as a black box function improving technique
as a whole. Furthermore, it has the ability to handle the microgrid working and
argument forecasting ambiguity.

Electricity load prediction is an important part of advanced power systems
i.e., SG, effective power controlling, and improved energy operation engineering.
Therefore, highly accurate prediction is needed for different perspectives, that
are related to control, forwarding, planning and unit responsibility in a grid.
Artificial Intelligence (AI) centered schemes has high competency to manipu-
late complicated mathematical problems, therefore, these techniques are widely
employed in number of research areas.
The Artificial Neural Network (ANN) outperforms statistical schemes, as

ANN is more efficient in mapping inputs to the outputs beyond complicated
mathematical designs. Diverse learning structures are used by ANN for exploit-
ing the linear association among the inputs [30]. ANN schemes has better depic-
tion than analytical and time series techniques for prediction problems. The
prediction performance in neural network is enhanced by the pre-processing of
training data, high equivalence impact, optimal network structure and better
learning algorithm. Moreover, ANN brings rapid confluence, minimized comput-
ing complexity, minimal training period and improved generalization [31].
Given a time series of 30 mins electricity loads, up to the time t, X1 ,.....Xt ,
our goal is to predict load at time t+1, i.e., Xt+1 .

Our proposed system forecasts the electricity load. In our proposed model we
used dataset from Australian Energy Market Operator (AEMO). In this dataset
electricity load recordings are taken after every 30 min, therefore producing 48
recordings for 48 time lags in a single day. Now if we want to predict the load of
a time lag, we have just 48 features to use. To increase the number of features
and better prediction we combine the records of week days from same week to
form a single record yielding 336 recordings for a single record, i.e we have 48 × 7
features. The proposed system model is visualized in Fig. 1.
Now we calculate the importance of each feature in updated dataset using
XGBoost feature selection technique. It help us to select the most appropriate
features for selection. We consider 35–40 features with highest importance values
for training and testing of our proposed scheme.
We have 1 year of data for 12 different regions, bringing 12 × 365 records for
365 days in a year. We divide the data into training and testing data as 75%
training and 25% testing data. We train our proposed model using training data
and perform testing using testing data. Algorithmic steps followed during the
process of load prediction using XGBoost are defined in Algorithm 1.

Algorithm 1. Proposed Scheme Algorithm for Load Prediction.

1: Load daily records data
2: Convert daily records data to weekly records data
3: for i ← 1 to size(f eatures) do
4: Calculate feature importance for feature
5: end for
6: Select features with importance value greater then threshold
7: Divide data into training and testing data
8: Train model over training data
9: Predict load using trained model over testing data
10: Calculate accuracy
11: Calculate MAPE
12: Calculate MAE
4.1 Evaluation Metrices
We used various standards for the evaluation of our proposed prediction model
efficiency. The two most commonly used metrices for the measurement of predic-
tion accuracy are Mean Absolute Percentage Error (MAPE) and Mean Absolute
Error (MAE).
4.1.1 MAPE
The MAPE may be a live of prediction accuracy of a forecasting methodology for
constructing fitted statistic values in statistics, specifically in trend estimation.
it always expresses accuracy as a proportion of the error. as a result of this range
may be a percentage, it may be easier to know than the opposite statistics. The
MAPE is outlined as shown in (1). Here, and area unit the actual worth and
therefore the forecast worth, severally. Also, is the number of times discovered.
n
100 At − Ft
M AP E = (1)
n t=1 At
4.1.2 MAE
In statistics, the MAE is employed to measure however shut forecasts or predic-
tions area unit to the particular outcomes. It’s calculated by making a mean of
absolutely the variations between the prediction values and therefore the actual
ascertained values. The MAE is defined as shown in (2). Wherever nine is that
the prediction price, is the actual price.
n n
1 1
M AE = |fi − yi | = |ei | (2)
n i=1 n i=1

We evaluated the performance of our proposed forecasting technique for predict-
ing the load. We did perform a lot of experiments. The results obtained from
extensive simulation are discussed here in this section.
5.1 Data Set Description

We used AEMO load data for the year 2017. The dataset records electricity load
after every 30 mins making 48 lags daily. The considered dataset has records
about 12 different areas for the same year 2017. We considered 365 days in a
year in the provided dataset. Here Fig. 2 plots the load profile for individual
week days. i.e., each weekday has its own line. From the graph we can see that
Saturday has an overall highest load with respect to the other week days in the
selected week. Furthermore, we can see that Wednesday has lowest load with
respect to the other weekdays in the selected week. Where Fig. 3 is displaying
the load of two consecutive weeks. It is clear from Figs. 2 and 3 that the electricity
load data has daily as well as weekly cycles. The load at a specific time with
respect to the other day is more or less same and it rises and falls more or less
like the previous and next day. Similarly for the week, we can see that the load
at one day in a week is more or less to the same day in previous and next week.
These cycles are due to the cycles in daily human activities. For performing the
load prediction of a specific time lag we combined the daily loads for same week
to form a weekly dataset. i.e We combined the data from same weekdays to form
single row. While conversion we neglected the weekdays that were not forming
a complete week at the end of the year. This conversion provides more features
for forecasting the load of a time lag.
Fig. 2. Daily load by weekdays

Fig. 3. Two weeks load
5.2 Prediction Model Configurations
We used XGBoost a gradient boosting framework, introduced back in 2014.

XGBoost can be used as a forecasting technique for feature selection and load
prediction of a time lag. From prediction to classification XGBoost has proved
its worth in terms of performance.
When we convert the dataset to form a weekly data we have 48×7 recordings
for a week. One row in the modified dataset represent a week with 336 features.
To understand the importance of features, we used XGBoost to calculate the
Feature Importance (FI) of all these features.
Figure 4 is depicting the FI of all these features. The greater the FI values
means the feature will more effect the load prediction. It is evident from Fig. 4
that features with hight importance values are less in number, where most of
the features have low importance value. We can see that the features close to
the predicting lag have high importance as well. Also it is evident that days
having same weekday for which we are predicting the load i.e., sunday have
high importance value. We will only use the features with high FI values for the
purpose of prediction and eliminate all other features from dataset. For selecting
the features, we set a threshold for feature importance and we set this threshold
by repeating experiment multiple times, and trying different threshold value.
The value with best results in considered as threshold. The Fig. 4 shows that the
features with high FI value are less in number, it will save the running time.
5.3 Forecasting Results
We used XGBoost for forecasting the load for a specific time lag in a week using
weekly data. Figure 5 shows the real load in the dataset and the XGBoost fore-
casted load. Here x-axis is depicting the time lags where y-axis is the load at
Fig. 4. XGB feature importance
that specific time lag. Actual load is represented by the blue graph, where fore
casted load is represented by the green graph. We can see that the XGBoost
load prediction follows the real load at most of the time, however at some high
load instances XGBoost is not exactly following the real load. We can see that
XGBoost is not predicting well at the high loads.
Fig. 5. XGB predictions
The XGBoost load forecasting results for a time lag are displayed in Fig. 6.
We can see that XGBoost forecasting technique results in a low Mean Average
Percentage Error (MAPE), high accuracy and high Mean Average Error (MAE).
XGBoost load prediction resulted in a 10.08% MAPE, 97.21% accuracy and
88.90% MAE.
Fig. 6. XGB results
6 Conclusion
In this paper, we proposed a new scheme for electricity load forecasting. We con-
verted daily electricity load information into weekly load information. It increases
number of features available for predicting load for a lag variable. Then, we used
XGBoost, a recently dominant machine learning technique for time series pre-
diction, for feature selection from converted data. Once features are extracted
we train the model using XGBoost. After training we use trained model for load
prediction.
References
1. Mathaba, T., Xia, X., Zhang, J.: Analysing the economic benefit of electricity price
forecast in industrial load scheduling. Electr. Power Syst. Res. 116, 158–165 (2014)
2. Sarada, K., Bapiraju, V.: Comparison of day-ahead price forecasting in energy
market using Neural Network and Genetic Algorithm. In: Proceeding of the Inter-
national Conference on Smart Electric Grid, pp. 1–5 (2014)
3. Shafie-Khah, M., Moghaddam, M.P., Sheikh-El-Eslami, M.: Price forecasting of
day-ahead electricity markets using a hybrid forecast method. Energy Convers.
Manag. 52(5), 2165–2169 (2011)
4. Garcia, R.C., Contreras, J., Van Akkeren, M., Garcia, J.B.C.: A GARCH forecast-
ing model to predict day-ahead electricity prices. IEEE Trans. Power Syst. 20(2),
867–874 (2005)
5. Shahidehpour, M., Yamin, H., Li, Z.: Market overview in electric power systems. In:
Market Operations in Electric Power Systems, pp. 1–20. Wiley, New York (2002)
6. Ci-wei, G., Bompard, E., Napoli, R., Cheng, H.: Price forecast in the competitive
electricity market by support vector machine. Phys. A: Stat. Mech. Appl. 382(1),
98–113 (2007)
7. Cai, Y., Lin, J., Wan, C., Song, Y.: A stochastic Bi-level trading model for an
active distribution company with distributed generation and interruptible loads.
IET Renew. Power Gener. 11(2), 278–288 (2017)
8. Weron, R.: Electricity price forecasting: a review of the state-of-the-art with a look
into the future. Int. J. Forecast. 30(4), 1030–1081 (2014)
9. Hu, L., Taylor, G.: A novel hybrid technique for short-term electricity price fore-
casting in UK electricity markets. J. Int. Counc. Electr. Eng. 4(2), 114–120 (2014)
10. Voronin, S., Partanen, J.: Forecasting electricity price and demand using a hybrid
approach based on wavelet transform, ARIMA and neural networks. Int. J. Energy
Res. 38(5), 626–637 (2014)
11. Kou, P., Liang, D., Gao, L., Lou, J.: Probabilistic electricity price forecasting with
variational heteroscedastic Gaussian process and active learning. Energy Convers.
Manag. 89, 298–308 (2015)
12. Shrivastava, N.A., Panigrahi, B.K.: A hybrid wavelet-ELM based short term price
forecasting for electricity markets. Int. J. Electr. Power Energy Syst. 55, 41–50
(2014)
13. He, K., Xu, Y., Zou, Y., Tang, L.: Electricity price forecasts using a curvelet
denoising based approach. Phys. A: Stat. Mech. Appl. 425, 1–9 (2015)
14. Wan, C., Xu, Z., Wang, Y., Dong, Z.Y., Wong, K.P.: A hybrid approach for prob-
abilistic forecasting of electricity price. IEEE Trans. Smart Grid 5(1), 463–470
(2014)
15. Wan, C., Niu, M., Song, Y., Xu, Z.: Pareto optimal prediction intervals of electricity
price. IEEE Trans. Power Syst. 32(1), 817–819 (2017)
16. Liu, B., Nowotarski, J., Hong, T., Weron, R.: Probabilistic load forecasting via
quantile regression averaging on sister forecasts. IEEE Trans. Smart Grid 8(2), 1
(2017). https://doi.org/10.1109/tsg.2015.2437877
17. Wanga, P., Liu, B., Hongb, T.: Electric load forecasting with recency effect: a big
data approach. Int. J. Forecast. 32, 585–597 (2016)
18. Chou, J.-S., Ngo, N.-T.: Smart grid data analytics framework for increasing energy
savings in residential buildings. Autom. Constr. 72, 247–257 (2016)
19. Ludwig, N., Feuerriegel, S., Neumann, D.: Putting big data analytics to work:
feature selection for forecasting electricity prices using the LASSO and random
forests. ISSN 1246-0125 (Print) 2116-7052
Energy 221, 386–405 (2018)
21. Wang, L., Zhang, Z., Chen, J.: Short-term electricity price forecasting with stacked
denoising autoencoders. IEEE Trans. Power Syst. 32(4), 2673–2681 (2017)
22. Lagoa, J., De Ridder, F., Vrancx, P., De Schutter, B.: Forecasting day-ahead elec-
tricity prices in Europe: the importance of considering market integration. Appl.
Energy 211(1), 890–903 (2018)
23. Raviv, E., Bouwman, K.E., van Dijk, D.: Forecasting day-ahead electricity prices:
utilizing hourly prices. Energy Econ. 50, 227–239 (2015)
24. Xiao, L., Jianzhou Wang, R., Hou, J.W.: A combined model based on data pre-
analysis and weight coefficients optimization for electrical load forecasting. Energy
82(15), 524–549 (2015)
and energy consumption forecasting. Energies 11, 452 (2018)
26. Moon, J., Kim, K.-H., Kim, Y., Hwang, E.: A short-term electric load forecasting
on Big Data and Smart Computing (2018)
27. González, J.P., San Roque, A.M., Perez, E.A.: Forecasting functional time series
28. Luo, J., Hong, T., Fang, S.-C.: Benchmarking robustness of load forecasting models
under data integrity attacks. Int. J. Forecast. 34, 89–104 (2018)
29. Dong, G., Chen, Z.: Data driven energy management in a home microgrid based
on Bayesian optimal algorithm. IEEE Trans. Ind. Inform
30. Fausett, L.: Fundamentals of Neural Networks: Architectures, Algorithms, and
Applications. Pearson Education, Delhi (2006)
31. Shekhar, S., Amin, M.B.: Generalization by neural networks. IEEE Trans. Knowl.
Data Eng. 4, 177–185 (1992)
Electricity Price Forecasting in Smart
Grid: A Novel E-CNN Model
Waleed Ahmad1 , Nadeem Javaid1(B) , Annas Chand2 ,

Syed Yousaf Raza Shah3 , Umar Yasin3 , Mudassar Khan4 , and Aimal Syeda1
1
2
COMSATS University Islamabad, Abbottabad Campus, Abbottabad, Pakistan
3
Riphah International University, Islamabad, Pakistan
4
Abasyn University, Islamabad, Pakistan
http://www.njavaid.com/
Abstract. The vital part of the smart grid is electricity price forecast-
ing because it makes grid cost saving. Although, existing systems for
price forecasting may be challenging to manage with enormous price
data in the grid. As repetition from the feature cannot be avoided and
an integrated system is needed for regulating the plans in price. To han-
dle this problem, a new price forecasting system is developed. This pro-
posed model particularly integrated with three systems. Initially, fea-
tures are selected from the random data by combining the Mutual Infor-
mation (MI) and Random Forest (RF). The Grey Correlation Analysis
(GCA) is used to remove the redundancy from the selected features.
Secondly, the Recursive Feature Elimination (RFE) scheme is used to
reduce the dimensions. Finally, classification is done based on Enhanced-
Convolutional Neural Network (E-CNN) classifier to forecast the price.
The simulation results show that our accuracy of the proposed system is
higher than existing benchmark schemes.
1 Background
The main objective of the smart grid is to reduce the energy peak load and
to balance the space between power supply and demand. Electricity consumers
can interact in the working of smart grid to conserve the energy and shift the
load. As for users, price forecasting is important because knowing the price they
can shift the load On and Off. In the last couple of years, high power demand
in domestic areas has been observed due to an expansion in electrical devices.
The power demand has extended up to 28% due to the rise in energy consump-
tion. Many researchers working on the residential area to control and manage
the household appliances. In [1], authors proposed a novel Memory Updation
Heuristic Scheme (MUHS), which efficiently schedule the appliances from on-
peak to off-peak time slots. We studied the literature work in price forecasting
and feature selection. Electricity price forecasting is done by time series pre-
diction and machine learning algorithms. In [2], authors proposed a model to
https://doi.org/10.1007/978-3-030-15035-8_109
Electricity Price Forecasting 1133
foretell the day ahead power price according to load demand and temperature,
with the usage of neural network design and evaluation of the unique spectrum.
The authors in [3], presented a probabilistic system to determine hourly price
rate, where the bootstrapping framework is used for examining uncertainty and
induced extreme learning machine strategy is introduced for wavelet neural net-
works. In [4], authors proposed a periodically Autoregressive Moving Average
Hilbertian (AMAH) design to determine the changeable average values in prac-
tical time series, which can be implemented to price prediction.
In [5], authors proposed a deep neural system model in the regional market to
increase the prediction accuracy, Furthermore, a new feature selection algorithm
is proposed to measure the feature importance by utilizing Bayesian optimiza-
tion. To improve accuracy, researchers forecast prices from two local markets
simultaneously. In paper [6], a novel price forecasting framework is proposed,
which consist of four deep learning models for price forecasting to increase pre-
diction accuracy. To examine the model extensively, the authors evaluated 27
algorithms and stated that our proposed deep model beats the benchmarks.
Prior researches mainly concentrate on feature algorithms or classifiers scheme,
where conventional classifiers, e.g., Decision Tree (DT) are used in [7].
Yet, DT normally faces the over fit obstacle, which indicates the DT shows
well in training but not in prediction. Luckily, with the advancement of artificial
intelligence, several artificial intelligence predictions algorithms, including Elman
Neural Network (ENN), Generalized Regression Neural Network (GRNN), Arti-
ficial Neural Networks (ANNs), Wavelet Neural Network (WNN) [8], etc., Some
classifiers like Extreme Learning Machine (ELM) [9] and Least Squares Support
Vector Machine (LSSVM) [10] has been improved in forecasting domains, such
as air conditions advanced warning [11]. However, the literature contains many
techniques for determining electricity price. Every technique has their own ben-
efits and limitations, hence cannot always accomplish the desired achievement.
Clearly, the analytical systems have a bad extrapolation result and tight fore-
casting range, with a considerable dependency on data. Furthermore, these are
proper for linear data and inappropriate for data featuring variations and noise.
The Grey Design (GD) is proper for data by having an exponential curve but
improper to follow the trend of a waved series [12]. From a common aspect,
numerous applications might start modifying 1-D Convolutional Neural Net-
work (CNN) for linear data used for time series forecasting [13]. Short term load
and price forecasting is also propose in [14,15]. A novel game theory approach
was also proposed for load forecasting in [16]. Short term load forecasting in
industrial area is also proposed in [17]. In this research work, an Enhanced-CNN
is proposed for price forecasting. We have applied tenser flow to develop our
system model.

Traditional grid system has a one-way communication channel and cannot com-
municate with end users. From the last decade, Researchers are working on
1134 W. Ahmad et al.
smart grid to improve its efficiency. The smart grid has a bidirectional digital
communication channel between end users and utility. The aim of the smart
grid is to reduce the cost and burden on utility. This can be done through price
forecasting. Several deep layer models are proposed in the field of time series
and energy-related sides to bring the best system model in the area of elec-
tricity price forecasting. Specifically, many neural networks models have been
presented which have limitations of hidden layer and accuracy e.g., Recurrent
Neural Network (RNN) [18] and Multilayer Perceptron (MLP) [19].
3 Proposed Solution
Based on above limitations, the main objective and focus of this paper are to
develop a deep layer model that can efficiently and accurately predict the next
day prices. In this paper a deep learning system model is developed which is
described as follows:
We presented an integrated price forecasting structure to obtain accurate big
data prediction in the smart grid. The primary effort in this paper is categorized
in hybrid feature selection, feature extraction, and classification based on E-CNN
are deployed in our system model.
To build this system model, a new Grey Correlation Analysis (GCA) feature
selection with Mutual Information (MI) and Random Forest (RF) are used to
measure the feature importance. In order to reduce the redundancy from the
selected features, we used Recursive Feature Elimination (RFE) algorithm. Fur-
thermore, we also design E-CNN to increase the accuracy than existing bench-
marks.
The efficiency of our system is evaluated through many simulations with ISO
New England real price data 2016. The results reveal that our system model
outperformed than benchmark schemes.
4 System Model
Our system model consists of three stages i.e., GCA based MI and RF feature
selection, RFE based feature extraction and E-CNN based classification.
4.1 System Model Objective

The main objective of our proposed model is to efficiently predict the electricity
price. To attain this objective, we need to analyze the price data carefully by
selecting the important features from the feature set and then enhance the clas-
sifier attentively. Hence, the following are some parameter which are important
in the improving of performance.
4.1.1 Classification Accuracy

This is the main objective of our designed model. We developed a new deep
layer E-CNN model and increase the classification accuracy from its benchmark
schemes.
4.1.2 Feature Reduction

In our model, hybrid feature selection affect the accuracy performance directly
to classification, so we choose MI and RF for careful feature selection.
4.1.3 Processing Time

The proposed model operated fast in price forecasting classification.
The major issues in electricity price prediction are time and accuracy per-
formance. However, there are many factors which influence the accuracy of elec-
tricity when training the data in the classifier. In order to improve the efficiency
of our proposed model, we make a hybrid feature selector, an RFE based feature
extraction and E-CNN based Classifier. The hybrid feature selector classifica-
tion model starts with the standardization of the input data. This is considered
as a difficult part because the whole system model depends upon the selected
standardized raw data as depicted in Fig. 1. In the next step, data fed into GCA
based hybrid feature selector, where MI and RF train the data in parallel. The
feature received by MI and RF are store in an index after analyzing the feature
by selector. Due to different objective of the selection scheme, this method could
be run separately. After that, RFE scheme is used to eliminate the redundant
data from the selected features. Our proposed system model completely depends
upon the redundant and importance features. Furthermore, the selected features
forwarded to SVM and E-CNN for classification. E-CNN is tuned by modifying
multiple layers in model.

5 Existing and Proposed Schemes

5.1 Decision Tree (DT)
Decision Tree takes the data and creates a sequence of conditions into a tree
formation. The nodes and roots of decision trees are comprised of attributes that
separate the data into various features according to conditions. After that classes
are named with node labels true or false. When the decision tree is developed,
the classification of a test set becomes easy. Beginning the node, we select the
relevant test set to record and follow the outcome. In order to foretell the class
label and to trace its way in a decision tree for the record, the route ended at a
leaf with no labeled.
5.2 Random Forest (RF)

RF classifier chooses the subset from the training set and generates a collection
of decision trees by random selection. Furthermore, aggregation is performed on
various DT to select the ultimate test class. RF creates a set of trees to enhance
the insufficient robustness and performance of DT. RF has only two parameter
for tuning: the first one is the set of trees and the second one is the variables
which split in every try.
5.3 Support Vector Machine (SVM)
SVM classifier is used to reduce the risk function. SVM super parameters and
risk function have a good relationship. Some SVM parameters are insensitive loss
function, cost penalty, and kernel. However, the tuning of these SVM parameters
is important and difficult task to achieve high accuracy and efficiency. The sim-
ple way to tune SVM parameters is with cross-validation and gradient descent.
Despite of these two tuning methods, computational complexity is increased
which reduce the performance of SVM.
5.4 Convolutional Neural Network (CNN)
CNN is a sample of network topology, generally practiced to use for deep learn-
ing systems. CNN is a controlled neural network. Similar to the conventional
structure of a neural network which consists of input, hidden and finally output
layers, the CNN also comprises of these features and input of one layer is the
output of another layer. Still CNN has some good features like pooling layer
and full connected layer. Of course, The CNN have more hidden layers than
conventional neural network.
6 Proposed Scheme
In this paper, we introduce a new approach by linking a deep learning design with
conventional machine learning systems to manage short term price forecasting.
Particularly, we utilize the E-CNN with clustering models and enhanced it to
obtain more accuracy in prediction. The conv1d function is used to feed the Con-
volutional layer with 1D data. As our model is linear so we used Relu activation
function to bypass the gradient vanishing problem.

0(x ≤ 0)
Relu(x) = (1)
x(x > 0)
Where x represents the convolution computation in layer and we selected the
0.1 value for tuning in our scenario. The process has two stages comprising of
four steps. The first step, we preprocess the input raw data of price to remove
redundancy from the features. In second step, we divide the whole raw data into
training and testing sets. In the third stage, an E-CNN is prepared with all of
these sets to develop the prediction model. The last stage validation is conducted
to evaluate the prediction model accuracy. The E-CNN further comprises of
two stages. In the first stage, the training data is preprocessed with traditional
feature selection and extraction algorithms and then E-CNN based system model
is constructed. In the second stage, testing data is taken as an input and ensemble
with E-CNN based system model to achieve better forecasting result. The four
layer E-CNN which comprise of Convolutional layer, activation function, max
pooling and fully connected layer showing in Fig. 2. Furthermore, by tuning all
these four layers we achieved the desired accuracy for price forecasting.
Fig. 2. Structure of E-CNN
6.1 E-CNN Algorithm
In E-CNN (Wi = 1, 2, ..L), represents the layer weights and n represents learning
rate. In this algorithm, randomly one weight is selected and reduce its learning
possibility to zero. When the training is over using back propagation and epocs,
the value of Wi is then modified to original n value. The details are described in
Algorithm 1. For the counter loop, the sub epocs is used which is set to zero in
nk. The predefined set values are Ci(i = 1, 2.., L). The benefit of using training
data is to train the neural network efficiency.
Algorithm for E-CNN
7 Simulation Setup
In order to examine the efficiency of our proposed model, we developed a python
framework according to the proposed system. Throughout the experiment, the
simulator is operating on HP system with Core i5, 6 GB RAM and 500 GB hard
disk. ISO New England Control Area (ISO NE-CA) 2016 [20] electricity price
information is used as the input for our proposed framework. The real-world
new England electricity price data consists of over 1000 records. The results of
the simulation are described as follows:
• 1. GCA based Hybrid feature selection,
• 2. RFE performance score,
• 3. E-CNN performance compared with benchmark schemes,
• 4. E-CNN robustness compared with benchmark schemes.

8.1 GCA Based Hybrid Feature Selection
GCA based hybrid feature selection is practiced on randomly selected features of

electricity price data. The ISO New England hourly electricity price data from
2016-1-1 to 2016-12-31 is taken as an input for simulation. In feature list, each
feature category has a pattern form as a vector. The components of this list depict
the feature values in various timestamps. Since our main focus is to predict the
electricity price, which is identified as the Regulation Clearing Rate (RegCR)
in the ISO data. Some features that impact the price slightly can be excluded.
Before the MI and RF hybrid feature selection, the correlation computation
between the various features is done with GCA as shown in Fig. 3. The various
grades of grey correlation are pointed, which shows that some features have
less grade than 0.5 are dropped, i.e., DA-CC, DA-MLC, RT-CC, and RT-MLC.
After the results obtained from GCA, the remaining features are selected for
evaluation. Through hybrid feature selection, the evaluation is performed to
obtain the importance of every feature. The feature selection results of MI and
RF are shown in Fig. 4. In order to increase the performance of the E-CNN
classifier, we have set some threshold T to select and drop the features. We
noticed that with rising in the threshold, features are discarded, which in turn
boost the training performance and the decrease the accuracy.
Fig. 3. Grey correlation grades depicting all features

Fig. 4. Feature importance achieved by MI and RF
8.2 Recursive Feature Elimination (RFE)
The aim of RFE is to choose the features by recursively analyzing the set of fea-
tures. In the beginning, the estimator is practiced on the original list of features
and the quality of an individual feature is acquired. Then, the small primary
features are extracted from the current list of features. This method is recur-
sively iterated on the clipped list until the required list of features to choose is
finally accomplished. To obtain the optimal figure of features, cross-validation
is employed with RFE to score various feature. The Fig. 5 presents a good RFE
cross-validated curve at 0.876 points and five important features are obtained.
This curve slowly declines in precision as the less important features are contin-
ued into the system model.
8.3 E-CNN Accuracy and Performance with Benchmark Algorithms
We evaluate the E-CNN with four benchmark classifiers i.e., RF, DT, SVM
and CNN. The Accuracy bars are shown in Fig. 6, which indicate that E-CNN
beats the benchmark schemes and achieved the accuracy of 97.03%. The E-
CNN curve line in Fig. 7 beats the actual and benchmark schemes by forecasting
less electricity price. The E-CNN shows satisfying results by determining less
electricity price. The performance among the benchmark algorithms indicates
that each module simulated in our framework can efficiently boost the accuracy.
The E-CNN efficiently decrease the inconsistency and repetition between the
features in order to increase the price forecasting accuracy.
Fig. 5. RFE feature score
Fig. 6. Accuracy comparison with benchmark algorithms

Fig. 7. Electricity price comparison
8.4 E-CNN Robustness Compared with Benchmark Algorithms
The robustness of a model is evaluated by two methods. The first method is to

include arbitrary noise (Error) to features and notice the average precision of
every algorithm. In the second method, random noise is added into some chosen
Fig. 8. Comparison on robustness

data which is not part of the same feature set. We selected the first method
and the solid lines in Fig. 8 showing the precision of algorithms. Basically, the
noise is produced by the failure of sensors. As the results showing that our
proposed E-CNN is more reliable as compared to other three existing algorithms.
In our proposed framework, each feature has a short impact of noise towards the
accuracy, because less important features are discarded initially in selection and
extraction phase from the feature set.
9 Conclusion
In this paper, a unique method to forecast electricity price by utilizing E-CNN

classifier and GCA based MI and RF feature selection and RFE based feature
extraction algorithm is proposed. Especially, GCA based MI and RF algorithm
is applied to sections the data into the feature set for building models lasting
and more efficient. This proposed system model is examined on a real-world
ISO new England price data. The experimental results evaluated by evaluation
metric Mean Absolute Percentage Error (MAPE) and by adding random noise
in the data, which reveal that the proposed approach beat the baseline models in
accuracy and robustness. For the future, we’ll additional quarry CNN forecasting
capabilities by adding additional similar elements in order to train the models.
Furthermore, we’ll investigate the outcomes of unique setting factors like Cluster
as well as iteration making CNN models.
References
1. Ahmad, W., Javaid, N., Sajjad, M.Z., Awan, T., Amir, M.: A new memory upda-
tion heuristic scheme for energy management in smart grid (2018)
2. Varshney, H., Sharma, A., Kumar, R.: A hybrid approach to price forecasting incor-
porating exogenous variables for a day ahead electricity market. In: IEEE Interna-
tional Conference on Power Electronics, Intelligent Control and Energy Systems
(ICPEICES), pp. 1–6. IEEE (2017)
3. Rafiei, M., Niknam, T., Khooban, M.-H.: Probabilistic forecasting of hourly elec-
tricity price by generalization of ELM for usage in improved wavelet neural net-
work. IEEE Trans. Ind. Inform. 13(1), 71–79 (2017)
4. González, J.P., San Roque, A.M., Perez, E.A.: Forecasting functional time series
5. Lago, J., De Ridder, F., Vrancx, P., De Schutter, B.: Forecasting day-ahead elec-
Energy 211, 890–903 (2018)
Energy 221, 386–405 (2018)
7. Wang, K., Yu, J., Yu, Y., Qian, Y., Zeng, D., Guo, S., Xiang, Y., Wu, J.: A survey
on energy internet: architecture, approach and emerging technologies. IEEE Syst.
J. PP(99), 1 (2017)
8. Wang, J., Yang, W., Du, P., Niu, T.: A novel hybrid forecasting system of wind
speed based on a newly developed multi-objective sine cosine algorithm. Energy
Convers. Manag. 163, 134–150 (2018)
9. Wang, J., Niu, T., Lu, H., Guo, Z., Yang, W., Du, P.: An analysis-forecast system
for uncertainty modeling of wind speed: a case study of large-scale wind farms.
Appl. Energy 211, 492–512 (2018)
10. Li, C., Xiao, Z., Xia, X., Zou, W., Zhang, C.: A hybrid model based on synchronous
optimisation for multi-step short-term wind speed forecasting. Appl. Energy 215,
131–144 (2018)
11. Xu, Y., Yang, W., Wang, J.: Air quality early-warning system for cities in China.
Atmos. Environ. 148, 239–257 (2017)
12. Yang, W., Wang, J., Wang, R.: Research and application of a novel hybrid model
based on data selection and artificial intelligence algorithm for short term load
forecasting. Entropy 19(2), 52 (2017)
14. Nayab, A., Javaid, N.: Load and price forecasting in smart grids using enhanced
support vector machine
(2018)
17. Wang, Z., Yan, W., Oates, T.: Time series classification from scratch with deep
neural networks: a strong baseline. In: 2017 International Joint Conference on
Neural Networks (IJCNN), pp. 1578–1585. IEEE (2017)
18. Goodfellow, I., Bengio, Y., Courville, A.: Deep Learning. MIT Press, Cambridge
(2016)
19. Panapakidis, I.P., Dagoumas, A.S.: Day-ahead electricity price forecasting via the
application of artificial neural network based models. Appl. Energy 172, 132–151
(2017)
20. ISO New England Energy Offer Data [Online] (2016). www.iso-ne.com/isoexpress/
web/reports/pricing/-/tree/dayahead-energy-offer-data
Electricity Price Prediction by Enhanced
Combination of Autoregression Moving
Average and Kernal Extreme
Learing Machine
Sahibzada Muhammad Shuja1 , Nadeem Javaid1(B) , Sajjad Khan1 ,

Umair Sarfraz1 , Syed Hamza Ali2 , Muhammad Taha2 , and Tahir Mehmood3
1
2
COMSATS University Islamabad, Wah Campus, Wah Cantonment 47040, Pakistan
3
Bahria University Islamabad, Islamabad 44000, Pakistan
Abstract. In electricity market, electricity price has some complicated

features like high volatility, non-linearity and non-stationarity that make
very difficult to predict the accurate price. However, it is necessary for
markets and companies to predict accurate electricity price. In this paper,
we enhanced the forecasting accuracy by combined approaches of Kernel
Extreme Learning Machine (KELM) and Autoregression Moving Aver-
age (ARMA) along with unique and enhanced features of both models.
Wavelet transform is applied on prices series to decompose them, after-
ward test has performed on decomposed series for providing stationary
series to AMRA-model and non-stationary series to KELM-model. At
the end series are tuned with our combine approach of enhanced price
prediction. The performance of our enhanced combined method is evalu-
ated by electricity price dataset of New South Wales (NSW), Australian
market. The simulation results show that combined method has more
accurate prediction than individual methods.
Keywords: Predictions of electricity price · ARMA · KELM ·

Wavelet transform · Enhanced combined price prediction
1 Background
In the past decade of electricity market, the volatile electricity prices became a
complicated phenomenon along with few characteristics and important concep-
tion. Market’s managers have to ensure the stability of forecasting for the power
market. In power, money contribution decisions and transmission inflation, pre-
dictions of electricity price play an important role. However, it is a complicated
task to predict accurate price of electricity because of many segregated fea-
tures, like as high volatility, nonlinearity, multiple seasonality, etc. In past few of
https://doi.org/10.1007/978-3-030-15035-8_110
researches, many authors have presented different prediction techniques to pre-

dict the price for electricity. Those techniques are majorly classified into three
categories, which are artificial intelligences, classical and data driven based pre-
diction techniques. With the intelligent feature and self-training process, neural
network as an artificial intelligence technique are best in performance over other
techniques. Some of the recent work is presented in [1].
In [2], forecasting techniques is summarized in the five different types: sta-
tistical models, multi-agent models, fundamental models, reduced-form models
and computational-intelligence model techniques. Statistical and computational
models are two widely used techniques among given five techniques. Statistical-
model techniques like Auto Regression (AR), Moving Average (MA), Autore-
gression Moving Average (ARMA) and Autoregression Integrated Moving Aver-
age (ARIMA), while some other ARMA-based models such as Exogenous vari-
able (ARMAX) and Generalization Autoregression Conditional Heteroskedastic
(GARCH) predict the current electricity price from historical data of price. How-
ever, statistical-model techniques are basically limited to capture the non-linear
and the rapid change behavior of electricity price signal [3].
The authors have proposed the recent learning technique in [4], which is
Extreme Learning Machine (ELM) relay on Single-hidden Layer Feed-forward
Neural-network (SLFN). This technique randomly produces the connection
weight among input, hidden layer and threshold of neurons in the hidden layer,
and they do not need the adjustment of parameters during the training process.
A framework for modeling for prediction of electricity price has been presented
in [5]. They proposed four different models of deep learning for electricity price
forecasting. The machine learning models outperform than statistical models.
However, their hybrid model was not capable to outperform in results.
Motivated by increasing the performance of forecasting in the market of elec-
tricity, two different methods of electricity price prediction are presented [6]. To
calculate the importance of features, a novel feature selection algorithm is used
for the prediction accuracy being improved from 15% to 12% Mean Average Per-
centage Error (MAPE). However, they have not expanded their experiment for
the European markets. With Singapores weekly forecasting price of electricity
is presented in [7]. They make use of ARIMA and complemented with GARCH
models for electricity price prediction. Results depict that their models consid-
erably match the price patterns based on out of univariate forecasted samples.
However, results conflict with multivariate data to be used in their models.
In [8], considering PJM and Spanish markets by authors to propose a com-
bination for hybrid method consist of current day-based forecaster, correlation
and Wavelet Transform (WT) analysis for preprocessing stage. Afterward, their
method results revealed positive prediction as compared with other respective
methods. An optimization scheme is proposed in [9], to minimize the price of
electricity. They have scheduled the usage pattern of residential appliances by
their scheme candidate solution updation algorithm. However, user comfort is
not considered to fulfill the consumer’s side waiting time for appliances usage in
the home.
Enhanced Combined Price Prediction 1147
Using publicly available data in [10], a forecasting method is proposed for

real time Ontario’s market of electricity. Their method uses data as input of
two different hourly based forecasted model. The model is capable to detect the
serious outrage and spikes in the electricity price. A model of functional additive
attributes for electricity price forecasting is presented in [11]. The proposed the
point wise prediction of information from the Spanish electricity market. Dataset
of the same paper is compared with other functional, regression models, which
makes better prediction result for load and price forecasting. A neural network
based training algorithm is presented in [12]. They have extended Kalman filter
and used energy price forecasted. One-step and n-step ahead of Kalman filter
verified better result of price prediction from European electricity system.
The authors presented a forecasting model for load predictions through
hybrid neural network [13]. In their work, they verified their simulations on
USA based datasets. The results of their method in comparison from two exist-
ing models outperform with respect of execution time, forecasting accuracy and
scalability. In [14], unique and new short-term forecasting method is presented
by using affinity propagation and a firefly heuristic algorithm for the system.
This technique improves various priority indexes to select similar days. Simula-
tions compared with the existing techniques for accuracy and operational cost,
which shows their system outperform and also minimized the error through error
measure.
Electricity price data in the real world have some unusual pattern, due to
which they are non-stationary and non-linear. Therefore, both of the non-linear
and linear prediction of electricity of price is needed for the effective purpose
of price prediction models. In this paper, proposed a combined price prediction
model have some enhanced feature of existing models KELM and ARMA. The
original series of electricity price is decomposed into more stable variants of
volatile and stationary series by WT. The ARMA-model is used for the predic-
tion of a stationary series because of its efficiency and robustness, whereas due to
best generalization of KELM-model performed forecasting on the volatile series
of price. To generate the final forecasted values of electricity price, proposed
method for model integrate the combination unique features from KELM and
ARMA. Dataset of New South Wales (NSW) is used to illustrate the forecasting
ability of our proposed method to compare the results with individual prediction
of KELM and ARMA on original price series.
In this section of the paper, we describe in detail proposed system model for
method develop to forecast the electricity price and is shown in Fig. 1. It is very
difficult to forecast the accurate electricity price, because of many circumstances
including weather, load and previous predicted prices. In addition, balance is
required between demand and supply as electricity cannot be stored due to the
inelastic nature of electricity shorter comes and unstable nature of generation
[15], complexity is produced in the behavior of electricity price. In this paper, we
capture different pattern of electricity price series to produce an accurate price

prediction. Generally, electricity price series presents irregular behavior due to
high fluctuation. In our proposed model, irregular pattern of price series are
change into a regular pattern of the essential series by WT and more accurate
results of prediction produced by decomposed series than direct price prediction
from forecasting model. However, wavelet decomposition series cannot reflect
the information received by individual model. Therefore, the performance might
not be fulfilled by single model prediction for WT as an input.

We used Enhanced Combined Prediction (ECP) of electricity price that can

perform on the basis of the unique characteristic of each model to get information
from decomposed price series. Our ECP method generates the better the predic-
tion output. In the ECP method, ARMA-model and KELM-model is used for
selection and training data. The ARMA-model is generally solved the problem
of linear time series; its performance for stationary series is quite satisfactory.
The positive feature of KELM-model is its fast speed of learning for non-linear
prediction. Hence, our new ECP method comprises of WT, ARMA and KELM
is proposed to forecast the price.
The performance steps in the ECP price model of prediction for electricity
price mentioned as follows:
• Preparation of dataset: Date of price is divided into training and testing

portion of the series. The data are trained for developing models and tested
for executing model.
• Applying wavelet transform: In this paper, Daubechine 5 is used as mother
wavelet and 5 levels of decomposition series. WT is used to decompose elec-
tricity price series into 5 levels, among them one is approximate series A and
others are detail series D1, D2, D3 and D4.
• Stationarity check of each series: We performed Augmented Dicky Fuller
(ADF)-test on decompose series for the check of stationarity. The ADF-test
results consist of ADF-statistic and three critical level values of 1%, 5% and
10%. If an ADF-statistic value less than critical level values than series are
stationary otherwise non-stationary.
• We used ARMA-model for processing of stationary series and KELM-model
for non-stationary series because of their construction.
• Combined and enhanced the forecasting of the individual result of KELM
and ARMA for the final prediction of electricity price.
2.1 Error Measures for Predictions
The performance ability of prediction of our proposed model is verified by dif-

ferent error measures. Most of the used measures are the Root Mean Square
Error (RMSE), Mean Square Error (MSE) and Mean Absolute Percentage Error
(MAPE). When error measures are computed at low values, its mean there is a
good performance of predictions. Lots of out range’s peaks for a price series of
electricity can numerously affect the prediction of price. In our proposed model,
MAPE outperform from other techniques ARMA and KELM.
3 Existing and Proposed Techniques
In this section, we discuss the techniques used for the predictions of electricity
price.
3.1 Kernel Extreme Learning Machine (KELM)
This model is used in many areas of researches with respect of simple ELM and
ELM comes out of SFLN. Its responsiveness behavior toward non-stationary and
volatile series is due to fast learning and forecasting ability in price. In ELM,
hidden layer’s parameters do not need tuning. In [17], Huang proposed the kernel
based ELM. Kernel matrix in ELM is used, when feature mapping is unknown
for the user. The user has not specified hidden nodes and the hidden layer feature
mapping. In the stable kernel function is changing the randomness of ELM that
why KELM is better in generalization.
Algorithm 1. Kernel Extreme Learning Machine

Require: Input: [Original Electricity Price Series;]
1: Train electricity series data;
2: Applying elmregressor to prepare model;
3: Applying fitter function;
4: Applying predict function;
3.2 Autoregression Moving Average (ARMA)
This is the model based on random time series used for predictive analysis [18].
Prediction analysis of time series values consist of three types: AR, MA and
ARMA. The ARMA model implementation is based on following steps:
• Adf-Test: for stationarity analysis.

• Identification of model feature and structure.
• Estimate the parameters and development of the model.
• Model testing for the noise.
• Predict the future values of cost.
Algorithm 2. Autoregression Moving Average

1: Train electricity series data;
2: Applying AR and MA coefficient parameters;
3: Combine both of parameters to prepare model;
4: Applying predict function;
3.3 Enhanced Combined Price Prediction (ECPP)
In the ECPP, to select a particular single model is an important part. The

ARMA-model is a quantitative model in which correlation function is used to
compare the data between current and historical activities. The ARMA-model
is generally used for prediction of stationary series. The dominance of KELM
is its capability for fast learning and generalizing, that’s why we use it for non-
stationary and volatile series. Performance of KELM-model is affected by penalty
factor and kernel parameters. With the help of few unique features from both
KELM and ARMA, we made our proposed technique of ECPP method.
Algorithm 3. Enhanced Combined Price Predictions

1: Applying Wavelet Transform to decompose series;
2: Two series Approximate and Detailed Series;
3: Applying stationarity adf-test on decomposed series;
4: Find local best solution;
5: if adf statitic < critical value then
6: Price Series= Stationary Series;
7: Applying ARMA on Price Series;
8: else
9: Price Series= Non Stationary Series;
10: Applying KELM on Price Series;
11: end if
12: Combine Output of both model;
13: Final Forecasting of Electricity Price;
4 Case Results and Reasoning

In this section of the paper, case of Australian Electricity Market (AEM) is
used to demonstrate the forecasting performance of the proposed method. Our
proposed method is tested on price prediction of AEM. We put a daily forecasted
price in our method from New South Wales (NSW) dataset for the month of
January of the year 2018. Each day of the month is used to verify the production
of our proposed method, although the training set of data comprised of 30 days
taken from the start of the test days. NSW dataset is taken from the website
[16].
4.1 Analysis of Case for Proposed Method
Applying wavelet decomposition, the two series are produced approximate-series

and detail-series. This series illustrates the low-frequency and high-frequency
components of the main series. We justify the simulation from our proposed
method in two steps.
First, stationarity analysis is performed through the ADF test in the wavelet
decomposition series of NSW data. The ADF test results show that the ADF-
statistic values of detail series, D1, D2, D3 and D4 are smaller than 1% critical
level, thus their unit root hypothesis is rejected. However, ADF-statistic value
of an approximate series A greater than 10% critical level, thus its unit root
hypothesis is not rejected. Stationarity analysis indicates that A is identified as
non-stationary series and other detail series are stationary series.
Second, we applied ARMA-model on stationary series contain four detail
series and KELM-model of non-stationary series having one approximate series.
After that, we enhanced the results for both of the models along with the sum
of their predicted price values.
4.2 Discussion and Comparison
To elaborate the efficiency of our proposed method for prediction, data from
NSW is tested for January month of the year 2018 and the three type errors
MAPE, MSE and RMSE are calculated. We compare the results of our enhanced
proposed scheme with ARMA-model and KELM-model for errors. NSW electric-
ity price forecasted data is predicted by ARMA-model is shown in Fig. 2. The
peak of predicted values depicts the accuracy of single ARMA-model.
The prediction of enhanced combined prediction for NSW data of electricity
price is depicted in the Fig. 4. Peaks of graphs show that when ARMA and KELM
for were used separately and were not combined, then the accuracy value price
prediction was less than the proposed scheme. Complex price feature cannot be
resolved by single KELM and ARMA. ARMA is not capable to solve the non-
linear issues, because its forecasted values are not more accurate than KELM.
In Fig. 3, single KELM-model based electricity forecasting of price is shown.
Fig. 2. ARMA-model base electricity price forecasting

Fig. 3. KELM-model base electricity price forecasting
MAPE error estimation is performed in Fig. 5. Our proposed enhanced

method is compared with other methods ARMA and KELM. The results show
that our proposed method has an efficient and better performance with respect
of other method. We have achieved significant reduction of error for MAPE by
8.67% in ARMA-model, 15.07% in KELM-model and 3.06% in our proposed
method.
Fig. 4. Enhanced combined electricity price forecasting
However, according to the analysis of our proposed method performance

is not satisfactory for error calculation by MSE and RMSE in Figs. 6 and 7.
Our proposed method achieves less accurate values for MSE and RMSE error
and encountered with more error rate as compared ARMA and KELM models
(Table 1).
Fig. 5. MAPE error in percentage comparison of our enhanced combined prediction

with ARMA and KELM
Table 1. Comparison of error measures
Methods(s) MAPE MSE RMSE

ARMA 8.67% 56.48% 56.48%
KELM 15.07% 5.36% 2.32%
ECPP 3.06% 11.44% 14.54%
Fig. 6. MSE error in percentage comparison of our enhanced combined prediction with
ARMA and KELM
Fig. 7. RMSE error in percentage comparison of our enhanced combined prediction

with ARMA and KELM
5 Conclusion
Due to uncontrolled electricity market, predicting the price of electricity has
become essential to power users and supplier. Implementing an efficient and an
accurate prediction model has become a very important role in the electricity
scope. In this paper, a new enhanced method for prediction of electricity price
combining KELM and ARMA forecasted output is proposed. Electricity dereg-
ulated price series is transformed and decomposed into different series by WT,
the KELM-model forecast the non-stationary part of decomposed price, while
ARMA-model is used to predict the stationary part of the price. Final forecasting
of electricity price is the combination of KELM-model and ARMA-model. NSW
market data is used to verify the forecasting ability of our proposed enhanced
combine method. This paper is completed depends on the improvement of fore-
casting accuracy through enhanced method. Although some prices patterns of a
specific time period are not matched, the proposed method improved the match-
ing feature for prediction of electricity price. In future, load and weather should
be considered for the prediction accuracy through our proposed method.
References
1. Cerjan, M., Krzelj, I., Vidak, M., Delimar, M.: A literature review with statistical
analysis of electricity price forecasting methods. In: 2013 IEEE EUROCON, pp.
756–763. IEEE (2013)
2. Weron, R.: Electricity price forecasting: a review of the state-of-the-art with a look
into the future. Int. J. Forecast. 30, 1030–1081 (2014)
3. Chan, S.-C., Tsui, K.M., Wu, H., Hou, Y., Wu, Y.-C., Wu, F.F.: Load/price fore-
casting and managing demand response for smart grids: methodologies and chal-
lenges. IEEE Signal Process. Mag. 29, 68–85 (2012)
4. Huang, G.-B., Zhu, Q.-Y., Siew, C.-K.: Extreme learning machine: a new learning
scheme of feedforward neural networks. In: 2004 Proceedings 2004 IEEE Interna-
tional Joint Conference on Neural Networks, pp. 985–990. IEEE (2004)
Energy 221, 386–405 (2018)
6. Lago, J., De Ridder, F., Vrancx, P., De Schutter, B.: Forecasting day-ahead elec-
Energy 211, 890–903 (2018)
7. Loi, T.S.A., Le Ng, J.: Anticipating electricity prices for future needs-Implications
for liberalised retail markets. Appl. Energy 212, 244–264 (2018)
8. Bento, P.M.R., Pombo, J.A.N., Calado, M.R.A., Mariano, S.J.P.S.: A bat opti-
mized neural network and wavelet transform approach for short-term price fore-
casting. Appl. Energy 210, 88–97 (2018)
9. Shuja, S.M., Javaid, N., Rafique, M.Z.: Towards Efficient Scheduling of Smart
Appliances for Energy Management by Candidate Solution Updation Algorithm
(CSUA) in Smart Grid
10. Chitsaz, H., Zamani-Dehkordi, P., Zareipour, H., Parikh, P.P.: Electricity price
forecasting for operational scheduling of behind-the-meter storage systems. IEEE
Trans. Smart Grid 9(6), 6612–6622 (2018)
11. Raa, P., Vilar, J., Aneiros, G.: On the use of functional additive models for elec-
tricity demand and price prediction. IEEE Access 6, 9603–9613 (2018)
12. Alanis, A.Y.: Electricity prices forecasting using artificial neural networks. IEEE
Lat. Am. Trans. 16(1), 105–111 (2018)
(2018)
15. Esther, B.P., Kumar, K.S.: A survey on residential demand side management archi-
tecture, approaches, optimization models and methods. Renew. Sustain. Energy
Rev. 59, 342–351 (2016)
16. AEMO. http://www.aemo.com.au/Electricity/Data/Price-and-Demand/Aggrega
ted-Price-and-Demand-Data-Files
17. Huang, G.-B., Chen, L.: Enhanced random search based incremental extreme learn-
ing machine. Neurocomputing 71, 34608 (2008)
18. Box, G.E., Jenkins, G.M., Reinsel, G.C.: Time Series Analysis: Forecasting and
Control. Wiley, Hoboken (2011)
Prediction of Building Energy
Consumption Using Enhance
Convolutional Neural Network
Hafiz Muhammad Faisal1 , Nadeem Javaid1(B) , Bakhtawar Sarfraz2 ,

Abdul Baqi2 , Muhammad Bilal2 , Inzamam Haider2 ,
and Sahibzada Muhammad Shuja1
1
Comsats University Islamabad, Islamabad 44000, Pakistan
2
Abstract. Electricity load forecasting plays a vital role in improving

the usage of energy through customers to make decisions efficiently. The
accuracy of load prediction is a challenging task because of random-
ness and noise disturbance. An extreme deep learning model is applied
in proposed system model to achieve better load prediction accuracy.
The proposed model used to extract features by combining the mutual
information (RF) and recursive feature elimination (RFE). Furthermore,
extreme learning machine (ELM) and enhance CNN are used for load
forecasting based on extracted features from MI and RFE. Addition-
ally, to check the performance of our proposed scheme, we compared it
with some benchmark schemes e.g. CNN, SVR and MLR. Simulation
results reveal that our proposed approach outperformed in prediction
performance.
Keywords: Load forecasting · Deep learning model ·

Mutual information and recursive feature elimination
1 Background
Rate of energy consumption is directly proportional with an increment in pop-

ulation. In energy consumption, an office building or building in a corporate
organization reserves a significant amount of energy [1,2]. In China, 2011, statis-
tical estimation in a report identifies that 28% of the total energy consumption is
account by the corporate buildings, which is going to reach 35% in 2020 [3]. Same
conditions are to be face in United States with energy consumption close to 39%.
Therefore, industry needs new innovative solutions in order to manage the energy
utilization. In this regard, construction of prediction models for building energy
consumption can play an important role. It can prove helpful in decision making
https://doi.org/10.1007/978-3-030-15035-8_111
Table 1. List of acronyms
ANN Artificial neural network

ARIMA Auto regressive integrated moving average
CART Classification and regression technique
CNN Convolutional neural network
DNN Deep neural network
KPCA Kernal principal component analysis
LSTM Long short term memory
MLP Multi layer perceptron
RFE Recursive feature elimination
RNN Recurrent neural network
ReLU Rectified linear unit
SG Smart grid
SVM Suport vector machine
MI Mutual Information
ELM Extreme learning machine
CDA conditional demand analysis
DL Deep learning
AI Artificial intelligence
MSPE Mean square percentage error
RMSE Root mean square error
MAPE Mean average percentage error
for building management to control the equipment in conducive way. A lot of

work has been done in data science. In the existing work, most commonly case
studies are dependent on the time series data of historical energy consumption
for prediction models construction [5]. Normally, energy consumption prediction
methods are divided into statistical and artificial intelligence (AI) methods. Sta-
tistical models for probabilistic models construction, utilize historical data future
energy utilization are estimated and analyzed on its basis. Author’s [6] proposed
principle component analysis for significant inputs selection in energy consump-
tion prediction model. Author’s [7] applies linear regression for electricity utiliza-
tion estimation while two comparative approaches; neural networks and fuzzy
modeling are utilized for linear regression performance evaluation. ARX extra
inputs in [8] was utilized in autoregressive model for building component param-
eters estimation. Similarly, Kimbara et al. in [9] constructs an autoregressive
integrated moving area (ARIMA) model for online building energy utilization
prediction construction. On the other side, ARIMAX model which is the ARIMA
representation with external inputs, is applied for power demand prediction of
the building. Author’s [11] proposed a method based on regression named as
Prediction of Building Energy Consumption 1159
conditional demand analysis (CDA) building energy consumption prediction. AI

and deep learning (DL) based models have the capability of prediction and the
empirical results obtained from these models are accurate more than applications
in real world. Therefore, these AI and DL-based models are being apply widely
in prediction for energy consumption for a building. Authors in research article
[12] proposed an innovative technique called clusterwise regression which results
in clustering and regression integration for foresees building energy consump-
tion. Author’s [13] proposed a clustering method. This method finds similarity
patters sequences which predicts electricity prices. Similarly in [14], k-means
methodology is proffered which analyzes electricity consumption patterns in a
building. On the other side in [15], a survey is done on managing electricity
related time series foreseeing using data mining techniques. To understand the
energy consumption patterns and levels, decision tree method is used in [16] and
[17] propose random forest algorithm to improve energy efficiency which helps
facility managers. Author’s [18] proposed a relevant data selection methodol-
ogy for consumption of energy prediction. As Artificial neural network system
(ANN’s) are also entrenched in prediction applications so [19] uses short-term
(ST) prediction model for electricity demands in a bioclimatic building. Author’s
[20] proposed two algorithms; Levenberg-Marquardt and OWO-Newton’s algo-
rithm for residential building energy consumption foreseeing (Table 1).
The remaining part of the paper is organized as follows. In Sect. 2, problem
statement is presented. In Sect. 3, system model is discussed in detail. The exper-
iments on the prediction of building energy consumption have been performed
in Sect. 5. Finally, the conclusions of the paper are shown in Sect. 6.
2 Problem Statement
In this paper, we focus the problem of electricity load forecasting. The aim of our
model is to predict the exact load forecast with maximum amount of data from
smart grid. To tackle load forecasting problem, four classifiers are applied to
predict the electricity load. CNN is a classifier that splits the data into training
and testing phase. CNN is efficient classifier, however, following problems are
need to answered for good accuracy of electricity load forecasting.
• Hard to tune parameters

• High computational complexity
3 System Model
Our proposed system model consist of four parts: normalizing, training testing
and prediction. Four classifiers with tuned parameters as shown in Fig. 1.
3.1 Preprocessing Data
System load data is collected for years 2015 to 2017 from ISO-NECA. Data is
divided on monthly basis and each years similar month data is grouped together
e.g. January 2015, January 2016 and January 2017 data is grouped together.
First three weeks of any month are used for the training process and the last
week is for testing. Data is normalized with maximum impact values and divided
into three major parts: training, testing and validation. To investigate the per-
formance of the four classifiers, we used three performance evaluation errors:
mean square percentage error (MSPE), root mean square error (RMSE) and the
mean absolute percentage error (MAPE) which can be calculated respectively as
TM
1
M SP E = |(Av − Fv )| (1)
T tm=1

TM
1
RM SE =
2
(Av − Fv ) (2)
T tm=1
TM
1 Av
M AP E = 100 (3)
T tm=1 Fv
where Av is the observed test value at time tm and Fv is the forecasted value
at time tm.
4 Proposed Scheme
In this section, we will discuss RFE and enhance CNN classifiers.
4.1 RFE
RFE method is used for feature selection. Feature selection method fits a model
and discards the weakest feature until the specified number of features is reached.
Algorithm 1. RFE Algorithm

1: Start
2: Train the model on the training set
3: for Each subset size A,i= 1 to A do do) do
4: Keep the record of A most important variables
5: data preprocessing
6: Determine the model performance
7: end for
8: End
4.2 Enhance CNN

In the domain of machine learning, CNN is a class of deep neural network. It
consists of more than one convolution layer and connected one by one. CNN has
one input layer, multiple hidden layers and one output layer. Hidden layers have
Convolutional layer, dense layer, flatten layer and max-pooling layer. Convolu-
tional layer achieves Convolutional operation and than input values passes to
the next layer. Feed forward neural network is used to train the network. max-
pooling is one of the pooling layer. The max pooling layer uses the maximum
number of neurons.

In this section, we are going to explain the following;
1. Simulation setup
2. Simulation results
3. Performance evaluation
Each of the item mentioned above are discuss in detail below:

To present our proposed model, we performed the simulation in python envi-
ronment. The simulation is performed on system with specification Core i7, 6th
generation, 640 GB hard disk and 8 GB RAM. Five years electricity load data
(ISO New England Control Area (ISO NECA) from 2010 to 2015) is used for the
evaluation of proposed scheme. Data is split into training and testing, in which
training and testing days are 822 and 274. The simulation results are organized
as follows
• Mutual Information (MI) and Recursive Feature Elimination (RFE) are used
for features selection
• Redundant features are discarded also with Recursive Feature Elimination

• In deep learning model, the autocorrelation function is used to calculate the
input values foe ELM
• Extreme Learning Machine (ELM) is applied as a predictor to obtain correct
prediction results
• ELM performance is measured with three evaluation metrics i.e. Mean Abso-
lute Percentage Error (MAPE), Mean Square Percentage Error (MSPE), Root
Mean Square Error (RMSE)
• The proposed scheme is compared with existing models, such as the Conven-
tional Neural Network (CNN), Recurrent Neural Network (RNN) Multiple
Linear Regression (MLR) and Support Vector Regression (SVR).

MI and RFE methods are applied to select the main features with respect to
target feature from electricity price data and electricity load. The features, which
have very low effect on price and load are discarded from data. In order to extract
the main feature MI is used as shown in Fig. 1. Figures 2 and 3 demonstrate that
the partial autocorrelation scheme is used, to find the input variables of the
extreme deep learning model (Figs. 4, 5, 6, 7, 8, 9 and 10).
Fig. 2. Feature selection

Fig. 3. Autocorrelation function
Fig. 4. Partial autocorrelation function

Fig. 5. Error rate
Fig. 6. CNN
Fig. 7. Enhance CNN

Fig. 8. Multi layer regression
Fig. 9. NN
Fig. 10. Support vector regression


For performance evaluation, four techniques are used i.e. MAE, MAPE, MSE,
and RMSE. RMSE has the lowest error values according to all mentioned classi-
fiers and MSE has the highest error value. In order to determined the accuracy
of the proposed model the comparison of performance evaluation is conducted
as shown in Fig. 11. Figure 11 shows that CNN has the highest error values so
proposed enhance CNN has low error values.
Fig. 11. Performance evaluation
6 Conclusion
Proposed deep learning shown its great learning prediction qualities. The
proposed model consists of two key parts; feature selection and classifiers. A
combination of two different techniques are used for feature selection. The pro-
posed model shows the best load prediction results in building sector. Pro-
posed model provided a novel prediction model for building energy consumption.
Comparison results indicate that the enhance CNN method performances better
than the existing machine learning methods.
References
1. Streimikiene, S.: Residential energy consumption trends, main drivers and policies
in Lithuania. Renew. Sustain. Energy Rev. 35, 285–293 (2014)
2. Ugursal, V.I.: Energy consumption, associated questions and some answers. Appl.
Energy 130, 783–792 (2014)
3. Hua, C., Lee, W.L., Wang, X.: Energy assessment of office buildings in China using
China building energy codes and LEED 2.2. Energy Build. 86, 514–524 (2015)
4. Zuo, J., Zhao, Z.Y.: Green building research-current status and future agenda: a
review. Renew. Sustain. Energy Rev. 30, 271–281 (2014)
5. Daut, M.A.M., Hassan, M.Y., Abdullah, H., Rahman, H.A., Abdullah, M.P.,
Hussin, F.: Building electrical energy consumption forecasting analysis using con-
ventional and artificial intelligence methods: a review. Renew. Sustain. Energy
Rev. 70, 1108–1118 (2017)
6. Li, K., Hu, C., Liu, G., Xue, W.: Building’s electricity consumption prediction using
optimized artificial neural networks and principal component analysis. Energy
Build. 108, 106–113 (2015)
7. Pombeiro, H., Santos, R., Carreira, P., Silva, C., Sousa, J.M.C.: Comparative
assessment of low-complexity models to predict electricity consumption in an insti-
tutional building: linear regression vs. fuzzy modeling vs. neural networks. Energy
Build. 146, 141–151 (2017)
8. Jimenez, M.J., Heras, M.R.: Application of multi-output ARX models for estima-
tion of the U and g values of building components in outdoor testing. Sol. Energy
79, 302–310 (2005)
9. Kimbara, A., Kurosu, S., Endo, R., Kamimura, K., Matsuba, T., Yamada, A.: On-
line prediction for load profile of an air-conditioning system. Ashrae Trans. 101,
198–207 (1995)
10. Newsham, G.R., Birt, B.J.: Building-level occupancy data to improve ARIMA-
based electricity use forecasts. In: Proceedings of the 2nd ACM Workshop on
Embedded Sensing Systems for Energy-Efficiency in Building, Zurich, Switzerland,
pp. 13–18, 2 November 2010
11. Aydinalp-Koksal, M., Ugursal, V.I.: Comparison of neural network, conditional
demand analysis, and engineering approaches for modeling end-use energy con-
sumption in the residential sector. Appl. Energy 85, 271–296 (2008)
12. Hsu, D.: Comparison of integrated clustering methods for accurate and stable
prediction of building energy consumption data. Appl. Energy 160, 153–163 (2015)
13. Alvarez, F.M., Troncoso, A., Riquelme, J.C., Ruiz, J.S.A.: Energy time series fore-
casting based on pattern sequence similarity. IEEE Trans. Knowl. Data Eng. 23,
1230–1243 (2011)
14. Pérez-Chacón, R., Talavera-Llames, R.L., Martinez-Alvarez, F., Troncoso, A.:
Finding electric energy consumption patterns in big time series data. In: Proceed-
ings of the 13th International Conference Distributed Computing and Artificial
Intelligence, Sevilla, Spain, 1–3 June 2016, pp. 231–238. Springer, Cham (2016)
15. Martı́nez-Álvarez, F., Troncoso, A., Asencio-Cortés, G., Riquelme, J.C.: A survey
on data mining techniques applied to electricity-related time series forecasting.
Energies 8, 13162–13193 (2015)
16. Tso, G.K.F., Yau, K.K.W.: Predicting electricity energy consumption: a compari-
son of regression analysis, decision tree and neural networks. Energy 32, 1761–1768
(2007)
17. Ahmad, M.W., Mourshed, M., Rezgui, Y.: Trees vs. neurons: comparison between
random forest and ANN for high-resolution prediction of building energy consump-
tion. Energy Build. 147, 77–89 (2017)
18. Paudel, S., Elmitri, M., Couturier, S., Nguyen, P.H., Kamphuis, R., Lacarriàre, B.,
Corre, O.L.: A relevant data selection method for energy consumption prediction of
low energy building based on support vector machine. Energy Build. 138, 240–256
(2017)
19. Mena, R., Rodrı́guez, F., Castilla, M., Arahal, M.R.: A prediction model based
on neural networks for the energy consumption of a bioclimatic building. Energy
Build. 82, 142–155 (2014)
20. Biswas, M.A.R., Robinson, M.D., Fumo, N.: Prediction of residential building
energy consumption: a neural network approach. Energy 117, 84–92 (2016)
21. Ahmad, A., Javaid, N., Mateen, A., Awais, M., Khan, Z.A.: Short-term load fore-
casting in smart grids: an intelligent modular approach. Energies 12(1), 164 (2019).
https://doi.org/10.3390/en12010164. ISSN: 1996-1073
22. Zahid, M., Ahmed, F., Javaid, N., Abbasi, R.A., Kazmi, S.Z., Javaid, A., Bilal,
M., Akbar, M., Ilahi, M.: Electricity price and load forecasting using enhanced
convolutional neural network and enhanced support vector regression in smart
grids. Electronics (2019). EISSN 2079-9292. (IF = 2.110, Q2)
Author Index
A Apostolou, Dimitris, 985

Aalsalem, Mohammed Y., 592 Arai, Yoshikazu, 537
Ababneh, Nedal, 946, 975 Asif, Rana Muhammad, 80
Abbas, Zeeshan, 1071 Ayub, Nasir, 582
Abbasi, Raheel Ahmad, 1097
Abbasi, Raza Abid, 80, 385, 1120 B
Abdellatif, Safa, 172 Balzano, Walter, 815, 822, 833, 841, 851, 860,
Abdullah, Muhammad, 67, 1071 870
Aburada, Kentaro, 683, 693 Baqi, Abdul, 1157
Abu-Salih, Bilal, 887 Barolli, Leonard, 205, 215, 234, 422, 604
Acheampong, Samuel, 25 Ben Hassine, Mohamed Ali, 172
Afanasyev, Ilya, 116 Ben Yahia, Sadok, 172
Afridi, Imran Uddin, 1107 Bilal, Muhammad, 1157
Ahmad, Waleed, 80, 385, 1132 Bouallegue, Ridha, 244
Aimal, Syeda, 582, 592 Boujnah, Noureddine, 244
Akmal, Hina, 398 Braytee, Ali, 923
Akther, Sumaiya Binte, 195 Bremie, Bushra, 887
Alahmari, Muteeb, 887 Butt, Suhail Ashfaq, 1084
Al-Ani, Ahmed, 923 Bylykbashi, Kevin, 205
Albahlal, Abdulaziz, 887
Albtoush, Teshreen, 887 C
Alhabashneh, Mohammad, 887 Calabrò, Stefano, 724
Ali, Muhammad, 385 Chan, Kit Yan, 887
Ali, Syed Hamza, 1145 Chand, Annas, 67, 1071, 1107, 1132
Aloisio, Alessandro, 734 Chang, Chuan-Yu, 786
Alqahtani, Abdullah, 887 Chang, Shih-Hao, 481
Alqahtani, Sulaiman, 887 Chang, Tsai-Rong, 793
Alshareef, Naser, 887 Changazi, Sabir Ali, 1084
Alsinglawi, Belal, 897, 907 Chen, Ding-Horng, 793
Alvi, Ashik Mostafa, 195 Chen, Haiyin, 757
Amanullah, 80, 1120 Chen, Hao, 757
Amato, Alessandra, 880 Chen, Hui-Tzu, 471
Amato, Flora, 815, 822 Chen, Jim Hao, 441
Ameri Sianaki, Omid, 915, 936, 946, 966, 975 Chen, Lin, 412
Anaissi, Ali, 923 Chen, Mei-Lan, 481

https://doi.org/10.1007/978-3-030-15035-8
1170 Author Index
Chen, Xiaolong, 757 Hakura, Jun, 537

Chen, Yan, 351 Hanif, Murtaza, 398
Chen, Yi-Chi, 451 Hasan, Md Anik, 195
Cheng, Xu, 3 Higashinaka, Naoki, 275
Cheraghian, Ali, 936 Higashino, Masayuki, 151
Chiang, Hsiao-Chi, 459 Honda, Junichi, 224
Chou, Ta-Yuan, 441 Hong, Jiangshui, 1055
Cozzolino, Giovanni, 815, 822, 880 Horn, Geir, 1041
Culmone, Rosario, 745 Hsieh, Meng-Ling, 451
Hsieh, Shan-Hsiung, 776
D Hu, Deng-Xiang, 768
Daneshgar, Farhad, 966 Hu, Jiaxi Alessia, 1055
Darwish, Omar, 897 Hu, Wu-Chih, 776
Das, Koushiq, 195 Huang, Pei-Yu, 776
Dawoud, Ahmed, 957 Huang, Ted, 481
de Luca, Alessandro, 822 Huang, Wun-Yuan, 441
de Oliveira Silva, Flávio, 183 Hussain, Md. Anwar, 702
de Souza Carvalho, William Chaves, 183 Hussain, Rasheed, 116
Dhoska, Klodian, 205 Hussain, Shafquat, 946, 975
Dimililer, Kamil, 713
Dreibholz, Thomas, 1007, 1055 I
Duan, Kevin, 887 Ikeda, Makoto, 215, 422
Duolikun, Dilawaer, 57, 570, 612 Inotani, Shimpei, 693
Durresi, Arjan, 641 Iqbal, Muhammad Mudassar, 1084, 1097
Iqbal, Zeshan, 371
E Ishida, Tomoyuki, 493, 516
Elhadad, Mohamed K., 334 Ishii, Naohiro, 44
Enokido, Tomoya, 57, 570, 612 Islam, Tahir, 592
Issa, Tomayess, 887
F Ito, Takanori, 524
Faisal, Hafiz Muhammad, 371, 1157 Iwamoto, Takumi, 266
Faiz, Md. Faiz Iqbal, 702 Izu, Tetsuya, 673
Fazalullah, Qazi, 398
Feng, Tianyi, 286 J
Feng, Yun, 633 Jan, Syed Qasim, 385
Fujio, Satoshi, 151 Jan, Tony, 915
Fujisaki, Kiyotaka, 266 Javaid, Nadeem, 67, 80, 103, 371, 385, 398,
Furukawa, Kazuyoshi, 673 582, 592, 1071, 1084, 1097, 1107,
1120, 1132, 1145, 1157
G Jiang, Ji-Han, 768
Gagliardi, Roberto, 724
Gebali, Fayez, 334 K
Gheisari, Soheila, 936 Kambara, Yusuke, 673
Ghuman, Muhammad Nauman Javid, 1120 Kao, Chen-Yu, 776
Gopu, Srinivas Reddy, 139 Karajeh, Ola, 897
Guo, Yinzhe, 612 Karim, Basit, 385
Gupta, Rahul Raj, 651 Katayama, Kengo, 362
Guruwacharya, Prabhat, 966 Katayama, Tetsuro, 693
Katayama, Yoshinori, 673
H Kaur, Davinder, 641
Haari, Keisuke, 275 Kennedy, Paul, 923
Habib, Shujaat, 371 Khacef, Kahina, 662
Haider, Inzamam, 1157 Khan, Abdul Basit Majeed, 1107
Hajati, Farshid, 936 Khan, Mudassar, 1132
Author Index 1171
Khan, Sajjad, 67, 80, 385, 398, 1107, 1145 Mohebi, Nicolay, 1020
Khan, Wahab, 1084 Moscato, Francesco, 815, 822
Khan, Wazir Zada, 592 Moscato, Vincenzo, 880
Khan, Zahoor Ali, 67, 1120 Mostarda, Leonardo, 734
Khan, Zain Ahmad, 398, 1071 Mubarak, Hasnain, 371
Kirsal Ever, Yoney, 713 Mubin, Omar, 907
Ko, Chien-Chuan, 768 Murakami, Shunsuke, 161
Kolici, Vladi, 205 Murano, Aniello, 870
Kuang, Hanbao, 1007
Kulla, Elis, 351
N
Kurita, Takeru, 604
Nabeta, Shinichi, 501
Nagatomo, Makoto, 683
L
Naji, Mohamad, 923
Lassoued, Narjes, 244
Nakamura, Shigenari, 570, 612
Lee, Hui-Lan, 441
Nakasaki, Shogo, 215
Lee, JooYoung, 116
Nandi, Sukumar, 129, 651
Li, Hui, 757
Nandi, Sunit Kumar, 129, 651
Li, Kin Fun, 334
Navarra, Alfredo, 734
Li, Man, 286
Naz, Aqdas, 1097
Li, Peng, 35, 561
Nung, Lin-Hui, 451
Liang, Xiaobing, 633
Lin, Chi-Yi, 471
O
Lin, Chun-Li, 776
Oda, Tetsuya, 362
Lin, Hsuan-Yu, 776
Odagiri, Kazuya, 44
Lin, Iuon-Chang, 459
Ogura, Naohiro, 57
Lin, Jung-Shan, 451
Ohmori, Motoyuki, 151
Lin, Pei-Shan, 793
Ohyanagi, Tatsuya, 516
Lin, Pei-Yu, 471
Oikawa, Takanori, 673
Lin, Wei-Tsen, 451
Okazaki, Naonobu, 683, 693
Lin, Yu-Xuan, 768
Oma, Ryuji, 612
Liu, Te-Lung, 441
Ozaki, Ryo, 362
Liu, Wen-chao, 93
Ozaki, Tomonobu, 305, 314
Liu, Yi, 234, 422
Ozera, Kosuke, 234
Liu, Yu-Jia, 14
Long, Wen-feng, 14
Luo, Yu, 1007 P
Pacini, Alessandro, 724
M Pagnotta, Fabio, 745
Maabreh, Majdi, 897 Pang, Jiayi, 561
Maeda, Hiroshi, 275 Papageorgiou, Nikos, 985
Maisto, Alessandro, 805 Park, Hyunhee, 256, 433
Majeed Khan, Abdul Basit, 1097 Park, Mirang, 683, 693
Mambretti, Joe, 441 Park, Seunghyun, 433
Mansourbeigi, Seyed M.-H., 623 Patiniotakis, Ioannis, 996
Marcantoni, Fausto, 724 Pelosi, Serena, 805
Marchang, Ningrinla, 702 Piermarteri, Andrea, 724
Materka, Katarzyna, 1041 Polito, Massimiliano, 805
Matsui, Kanae, 295 Pradhan, Dipesh, 1030
Matsuo, Keita, 422, 604 Prakash, Satya, 129
Mazzara, Manuel, 116 Przeździȩk, Tomasz, 1041
Mehmood, Tahir, 1145 Pujolle, Guy, 662
Meng, Xiang Zheng, 275
Mentzas, Gregoris, 985, 996 Q
Micheletti, Matteo, 724 Qasim, Umar, 371
Minieri, Maurizio, 860 Qureshi, Talha Naeem, 1084
1172 Author Index
R Tasneem, Nazifa, 195

Rashid, Fahad, 1107 Torii, Satoru, 673
Raun, Chun, 957 Tsagkaropoulos, Andreas, 985
Reddy, Rajidi Satish Chandra, 139 Tsai, Lung-Yu, 786
Rehman, Amjad, 582 Tsukiji, Hayato, 324
Ren, Haiying, 286 Tsuruoka, Toru, 161
Rivera, Victor, 116
Rosa, Pedro Frosi, 183 U
Uchida, Noriki, 493, 510, 516, 545, 550
S Uddin, Md Nasir, 25
Saito, Takumi, 570 Ueda, Chiaki, 362
Saito, Yoshia, 537 ur Rehman Hashmi, Muhammad Aqeel, 1097
Sajjad, Hassan, 592 Ur Rehman, Shujat, 1120
Sakamoto, Takuma, 161 ur Rehman, Shujat, 80
Sakuraba, Akira, 524 Uslu, Suleyman, 641
Salman, Hasan, 25 Usuzaki, Shotaro, 693
Sarfraz, Bakhtawar, 1157
Sarfraz, Umair, 1145 V
Sato, Goshi, 510, 545, 550 Verginadis, Yiannis, 985, 996
Sato, Kazuki, 314
Saurabh, 129 W
Schenkel, Joseph Adam, 1055 Wang, Tzu-Ying, 793
Sekeroglu, Boran, 713 Waqas, Muhammad, 1071
Shah, Syed Yousaf Raza, 1132 Wei, Bin, 93, 412
Shahristani, Seyed, 957 Wongthongtham, Pornpit, 887
Shen, Leixian, 561 Wu, Chien-Chung, 768
Shen, Rong-Show, 793 Wu, Di, 286
Shibata, Yoshitaka, 493, 510, 516, 524, 537,
545, 550
Shimizu, Shogo, 44 X
Shuja, Sahibzada Muhammad, 67, 398, 1145, Xu, He, 25, 35, 561
1157 Xu, Meng, 633
Singh, Pranav Kumar, 129, 651 Xue, Donghui, 35
Skrzypek, Paweł, 1041
Song, Tao, 757 Y
Sorrentino, Loredana, 870 Yamaba, Hisaaki, 683, 693
Spaho, Evjola, 205 Yang, Hai-bin, 93, 412
Stingo, Michele, 805 Yao, Wu-jun, 93, 412
Stranieri, Silvia, 833, 841, 851, 860, 870 Yasin, Umar, 1132
Strugar, Dragos, 116 Yeh, Fei, 441
Su, Yi-Jen, 776 Yokoyama, Hisashi, 224
Sultana, Tanzeela, 582 Yoshino, Yu, 215
Sun, Ningping, 161 Yousefi, Ashkan, 915
Sun, Wei-dong, 14 Yu, Suhan, 451
Syeda, Aimal, 1132 Yuan, Mengxin, 3
Yuze, Hiroaki, 501
T
Taha, Muhammad, 1145 Z
Tahir, Aroosa, 582 Zahid, Feroz, 1020, 1030
Tajima, Hirohisa, 224 Zeinali, Behnam, 936
Takano, Kosuke, 324 Zhai, Feng, 633
Takatsuka, Kayoko, 693 Zhang, Jing, 3
Takizawa, Makoto, 44, 57, 205, 234, 422, 570, Zhang, Qingyun, 561
612 Zheng, Li-ming, 14
Tan, Xiaodong, 3 Zhou, Xing, 1007

Web, Artificial Intelligence and Network Applications

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Web, Artificial Intelligence and Network Applications

Uploaded by

Copyright:

Available Formats

Advances in Intelligent Systems and Computing 927

More information about this series at http://www.springer.com/series/11156

Fatos Xhafa Tomoya Enokido

Web, Artiﬁcial Intelligence

Fatos Xhafa Tomoya Enokido

ISSN 2194-5357 ISSN 2194-5365 (electronic)

Library of Congress Control Number: 2019933333

© Springer Nature Switzerland AG 2019

Welcome to the 33rd International Conference on Advanced Information

Finally, we would like to thank Matsue City, Shimane Prefecture, The

Welcome to the 33rd International Conference on Advanced Information

Welcome to AINA-2019 Workshops to be held in conjunction with the 33rd

AINA-2019 Organizing Committee

Tomoya Enokido Rissho University, Japan

Program Committee Co-chairs

Akimitsu Kanzaki Shimane University, Japan

Hui-Huang Hsu Tamkang University, Taiwan

International Special Issue Journal Co-chairs

Fatos Xhafa Technical University of Catalonia, Spain

Marek Ogiela AGH University of Science and Technology, Poland

Arjan Durresi IUPUI, USA

International Liaison Co-chairs

Nadeem Javaid COMSATS University Islamabad, Pakistan

Local Arrangement Co-chairs

Elis Kulla Okayama University of Science, Japan

Makoto Ikeda Fukuoka Institute of Technology, Japan

Donald Elmazi Fukuoka Institute of Technology, Japan

Steering Committee Chairs

Makoto Takizawa Hosei University, Japan

Welcome to the 15th International Symposium on Frontiers of Information Systems

We wish all of you entertaining and rewarding experience in FINA-2019

FINA-2019 Organizing Committee

FINA-2019 Symposium Co-chairs

Akimitsu Kanzaki Shimane University, Japan

Program Committee Members

Leonard Barolli Fukuoka Institute of Technology, Japan

Welcome to the 15th International Workshop on Heterogeneous Wireless Networks

HWISE-2019 Organizing Committee

Leonard Barolli Fukuoka Institute of Technology, Japan

Makoto Takizawa Hosei University, Japan

Program Committee Members

Makoto Ikeda Fukuoka Institute of Technology, Japan

Web Administrator Co-chairs

Donald Elmazi Fukuoka Institute of Technology, Japan

Welcome to the 12th International Symposium on Mining and Web (MAW-2019),

MAW-2019 Organizing Committee

Kin Fun Li University of Victoria, Canada

Program Committee Members

Takahiro Hara Osaka University, Japan

Welcome to the 12th International Workshop on Bio and Intelligent Computing

BICom-2019 Organizing Committee

Leonard Barolli Fukuoka Institute of Technology, Japan

Makoto Takizawa Hosei University, Japan

Program Committee Members

Mimoza Durresi European University of Tirana, Albania

Web Administrator Co-chairs

Kevin Bylykbashi Fukuoka Institute of Technology, Japan

We wish to extend a warm welcome to everyone attending the 13th International

TeNAS-2019 Organizing Committee

Steering Committee Chair

Timothy K. Shih National Central University, Taiwan

Rubem Pereira Liverpool John Moores University, UK

Teh-Sheng Huang ChungHwa Telecom, Taiwan